Online Appendixes Ex-Post Merger Evaluation in the UK Retail Market for Books
A Service of
Leibniz-Informationszentrum econstor Wirtschaft Leibniz Information Centre Make Your Publications Visible. zbw for Economics
Aguzzoni, Luca; Argentesi, Elena; Ciari, Lorenzo; Duso, Tomaso; Tognoni, Massimo
Article — Accepted Manuscript (Postprint) Ex-post Merger Evaluation in the U.K. Retail Market for Books
The Journal of Industrial Economics
Provided in Cooperation with: German Institute for Economic Research (DIW Berlin)
Suggested Citation: Aguzzoni, Luca; Argentesi, Elena; Ciari, Lorenzo; Duso, Tomaso; Tognoni, Massimo (2016) : Ex-post Merger Evaluation in the U.K. Retail Market for Books, The Journal of Industrial Economics, ISSN 1467-6451, Wiley, Hoboken, Vol. 64, Iss. 1, pp. 170-200, http://dx.doi.org/10.1111/joie.12099
This Version is available at: http://hdl.handle.net/10419/180821
Standard-Nutzungsbedingungen: Terms of use:
Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Documents in EconStor may be saved and copied for your Zwecken und zum Privatgebrauch gespeichert und kopiert werden. personal and scholarly purposes.
Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle You are not to copy documents for public or commercial Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich purposes, to exhibit the documents publicly, to make them machen, vertreiben oder anderweitig nutzen. publicly available on the internet, or to distribute or otherwise use the documents in public. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, If the documents have been made available under an Open gelten abweichend von diesen Nutzungsbedingungen die in der dort Content Licence (especially Creative Commons Licences), you genannten Lizenz gewährten Nutzungsrechte. may exercise further usage rights as specified in the indicated licence. www.econstor.eu This is the peer reviewed version of the following article: Luca Aguzzoni, Elena Argentesi, Lorenzo Ciari, Tomaso Duso, Massimo Tognoni: Ex Post Merger Evaluation in the U.K. Retail Market for Books. In: The Journal of Industrial Economics 64 (2016), 1, S. 170-200, which has been published in final form at http://dx.doi.org/10.1111/joie.12099. This article may be used for non-commercial purposes in accordance with Wiley Terms and Conditions for Self-Archiving.
Ex-post Merger Evaluation in the UK Retail Market for Books
Luca Aguzzoni Chief Economist Team, European Commission Elena Argentesi University of Bologna Lorenzo Ciari European Bank for Reconstruction and Development Tomaso Duso Deutsches Institut für Wirtschaftsforschung (DIW Berlin) and Düsseldorf Institute for Competition Economics (DICE) Massimo Tognoni UK Competition & Markets Authority
September 19, 2014 Abstract: This paper evaluates the price effects of the merger of two major UK book retailers. We use a dataset containing monthly scanner data on a sample of 200 books in 50 local markets for four years around the merger. We compare the price changes after the merger in shops located in areas where both chains were present before the merger and in areas where only one chain was present. We also investigate the country-wide effect of the merger. We find that the merger did not result in any price increase either at the local or at the national level.
Keywords: Mergers, Ex-post Evaluation, Book market, Retail sector
Corresponding author: Elena Argentesi, Department of Economics, University of Bologna, Piazza Scaravilli 2, 40126 Bologna, Italy, Tel: + 39 051 2098661, Fax: +39 051 2098040, E-Mail: [email protected]. We wish to thank the Editor and two anonymous referees for very valuable suggestions. This paper is partially based on a research project we undertook for the UK Competition Commission (CC). Paolo Buccirossi and Cristiana Vitale were coauthors of that report and offered regular guidance for this paper; therefore we extend our sincere gratitude for their advice and contributions. We are also grateful to Luca Barbarito, Claire Chambolle, Amrita Ray Chaudhuri, Peter Haan, Cristian Huse, Adam Lederer, Chiara Monfardini, Massimo Motta, and Daniel Rubinfeld as well as our audiences at various seminars and conferences for very useful discussions and suggestions. During the course of this study, the CC staff provided invaluable support for which we are grateful. We also thank Giulio Altomari, Roberto Alimonti, Roberto Cervone, and Carlotta Dandolo for their excellent research assistance.
1 1. Introduction
There is a growing interest in retrospective merger studies, which comes principally from the need to understand how mergers alter market structure and welfare. Yet, additional motivation behind this increased interest lies in the perceived need for antitrust agencies to check and improve the effectiveness of their decision making (e.g. Kovacic, 2009). Moreover, from an academic perspective, retrospective merger evaluations are seen as useful in validating structural models for merger simulation, which are increasingly used as an ex-ante instrument to assess policy changes.1 Despite a growing number of studies analyzing the price effect of mergers in a variety of industries, little work exists on the ex-post evaluation of mergers in the retailing sector.2 This is particularly surprising since not only do retail markets constitute a significant part of the economy in terms of value added and employment in all developed countries but these markets are experiencing a great deal of merger activity. Moreover, as noted by Hosken et al. (2012), mergers in retailing are often subject to antitrust scrutiny. For instance, out of 176 grocery markets subject to merger control by the US Federal Trade Commission between 1996 and 2011, 152 were challenged.3 Similarly, the European Commission (EC) reviewed an increasing number of retailing mergers over the past two decades. Out of the 167 mergers in retail trade analyzed by the EC between 1990 and 2008, 13 were either blocked or approved with specific conditions and obligations. Mergers in retailing sectors present some specific features that differentiate them from concentrations in other markets, features that should not just be considered in the merger review process but also in any retrospective study. Indeed, the opening passage of the 2011 joint UK Competition Commission/Office of Fair Trading report on retail mergers (Competition Commission and Office of Fair Trading, 2011; p. 4) reads, “Retail
1 See Nevo and Whinston (2010) and Weinberg (2011). For a contribution that uses ex-post merger analysis to validate a simulation model, see Björnerstedt and Verboven (2012). 2 Among these very few papers, Hastings (2004) analyzes the effects of a merger in the gasoline retail market in the US. Skrainka (2012) studies instead the effects of the merger between two UK grocery retailers using consumer data. See both Ashenfelter, Hosken and Weiberg (2009) and Duso (2012) for reviews of the literature on ex-post merger studies. 3 See Horizontal Merger Investigation Data, Fiscal years 1996 – 2011, Federal Trade Commission Table 4.2. Available at: http://www.ftc.gov/os/2013/01/130104horizontalmergerreport.pdf.
2 mergers account for a significant number of cases that come before the Office of Fair Trading and the Competition Commission…. Moreover, some of the questions that such mergers raise are largely specific to the sector.” In particular, they are characterized by dispersed buyers and sellers (Davis, 2006): Consumers tend to make their purchases within their local shopping location and retail businesses generally have multiple outlets across a country. Yet, retail offers may be set either nationally or locally, which creates interplay between local and national competition. This aspect is particularly relevant for merger assessment, since it implies that any concentration may impact competition at different geographical levels. The focus of the few existing academic papers and most of the policy studies on retailing industries is on mergers in the food sector. However, other retail markets are also important, such as the book market, which shares several peculiar features with the markets for creative goods such as music records, movies, video games, and software. Given the unpredictable and usually short lasting popularity, books are generally characterized by a very short life cycle as well as by uncertain demand and short periods of high profitability.4 These features pose particular challenges for the analysis of the effect of competition on prices in this industry. The same book title may have a different value to consumers in different periods of time and, therefore, retailers’ pricing policy may change accordingly. Moreover, publishers have a rapidly changing portfolio of books. Therefore, it would be incorrect to identify the effect of the merger by simply looking at the price evolution of a constant sample of products. The above considerations suggest the need of expanding the existing evaluation methodologies and the importance of providing new empirical evidence to improve our understanding of competition in retailing sectors and, specifically, cultural goods markets. This study tries to address these issues by analyzing the effect of a consummated merger in the book retail sector. Waterstone’s acquisition of Ottakar’s –two of the major book retailers in the UK at that time– was cleared by the UK Competition Commission (CC) in 2006. During its investigation, the CC identified problematic local markets where both the merging parties competed head-to-head before the merger (overlap areas) and
4 See Canoy et al. (2006) and Beck (2007) for in-depth descriptions of the features of the book industry.
3 where the concentration was expected to have the potential for significant anti- competitive effects. Hence, to identify the causal effect of this merger, we compare the price evolution of the products sold by the merging parties in the overlap areas with the price evolution of their products in areas where only one chain was present pre-merger (non-overlap areas) by employing a differences-in-differences (DiD) approach.5 We quantify the price effects of the merger by using a rich dataset containing scanner data information on a sample of 200 books, sold in 60 stores, across 50 different local markets over a period of four years around the merger (2004-2007). As outlined in a joint report by the CC and the Office of Fair Trading (Competition Commission and Office of Fair Trading, 2011), it is often difficult to find evidence of local effects of mergers because of data limitations. On this issue, our case study has a major advantage with respect to previous studies since we have accurate data at different levels of aggregation, which allow us to assess both the local and the national effect of the concentration. Moreover, the peculiar structure of our database helps us alleviate the main potential shortcoming of the DiD methodology: the choice of the appropriate control group. As discussed by Nevo and Whinston (2010), finding a suitable control group to estimate the causal effect of a merger on prices is difficult in many industries. In this respect, the features of the retailing sector make it a good field of application, since one can exploit the variation in the competitive conditions across local markets (e.g., Hastings, 2004 and Choné and Linnemer, 2012). More specifically, we make an accurate choice of counterfactual areas by using a methodology that allows us to select non-overlap areas that closely match overlap areas in terms of observable demand and supply characteristics. In addition to analyzing the effects of the merger on local pricing, the structure of our dataset allows us to also investigate the country-wide effect of the merger on prices. The national analysis provides important complementary evidence on the effect of the
5 Other papers pursing the DiD methodology in merger analysis include Focarelli and Panetta (2003); Hastings (2004); Chandra and Collard-Wexler (2009); Ashenfelter and Hosken (2011); Ashenfelter, Hosken and Weinberg (2013); Choné and Linnemer (2012). See Weinberg (2008) and Hunter, Leonard and Olley (2008) for a survey. A growing strand of literature instead follows a structural approach; see Friberg and Romahn (2012), Skrainka (2012), and Björnerstedt and Verboven (2012).
4 merger by using a similar empirical framework as for the local analysis, but with a different level of data aggregation and different counterfactuals. At the local level, the results of our average treatment effect analysis show that there is no significant difference in prices after the merger between non-overlap and overlap areas where the merger should have been reasonably expected to generate the strongest effect. Yet, to more precisely identify the impact of the merger on the price of differentiated products, we further estimate heterogeneous treatment effects by exploiting the richness of our data along different dimensions: firm-specific, book-specific, and market-specific. First, the analysis of firm-specific variables suggests that there was a price convergence among the merging parties in overlap areas after the merger for most book categories. Second, book-specific heterogeneous treatment effects show some evidence of a change of the timing of discounts for products introduced after the merger, as suggested by Ashenfelter et al. (2013). Third, market-specific variables that should capture the intensity of local competition do not seem to play a significant role in explaining the causal effect of the merger. Consistently with the local analysis, we do not find any significant price increases due to the merger at the national level either. Finally, we also consider the possibility that the merger generates non-price effects. In particular, we provide indirect evidence of the fact that the merger did not negatively affect the quality of service or product variety by analyzing the impact of the merger on the merging parties’ sales volume. Our findings of no average effect of the merger on prices are in line with the considerations that the CC expressed in its decision. In particular, the CC deemed unlikely that the merger would lead to an increase in price in overlap areas with respect to non-overlap areas, due to the fact that the merged entity “would continue to face competition from, among others, WH Smith, the supermarkets, and the Internet retailers.” (HMV Group plc and Ottakar’s plc - Proposed acquisition of Ottakar’s plc by HMV Group plc through Waterstone’s Booksellers Ltd”, 12th of May 2006; p. 36).Similarly, the CC did not foresee any significant effect of the merger on national prices, neither for best-sellers, on which the parties face strong competition from supermarkets and other
5 low-priced retailers, nor for deep-range titles, on which they face increasing competition from Amazon and other online retailers. The paper is structured as follows. In the next Section we discuss the institutional setting and, in particular, the characteristics of the book industry and the merger. Section 3 contains a description of our dataset. We then present our main empirical analysis in Section 4 and the analysis on the effect of the merger at the national level in Section 5. Section 6 concludes.
2. The Book Industry and the Merger
The supply chain of the book industry is characterized by three groups of players: publishers, wholesalers, and retailers. Our analysis focuses on the latter. Retailers can be broadly classified into four groups: 1) retailers specialized in the sales of books, as well as small independent bookshops; 2) non-specialist retailers for which books are an important category; 3) retailers for which books are part of a wide range of goods, such as supermarkets and major multiples; and 4) online book retailers. These categories differ in the range of titles they hold: specialist shops and online retailers offer a large selection, whilst supermarkets and major multiples hold fewer titles, mainly current best-sellers. In the UK retail book market, as well as in other countries, pricing takes the form of setting the level of the discount off the recommended retail price (RRP), which is usually printed on the book by the publishers and acts as a ceiling for the retail price.6 Publishers generally set the RRPs according to estimates on what the market would bear and to cost-related demand shifters (type of binding, presence of colored images, etc.). Clerides (2002) provides evidence of the fact that book prices seem to depend more on cost-related demand shifters than on pure demand shifters (new editions, author’s previous publications). Beck (2004) empirically analyzes the role of resale price maintenance in the book industry. In the UK market, discounts differ by title categories, which are defined on the basis of the sales’ ranking. In particular, discounts are generally larger for bestsellers –i.e. the top 5,000 titles sold in one year– than for deep-range titles – i.e. the remaining titles. Indeed, the cut-off of 5,000 was identified by the CC as the
6 Davies et al. (2004) and a report by the Office of Fair Trading (Office of Fair Trading, 2008) also provide in-depth overviews of pricing policies and regulations in the UK book market.
6 threshold separating these two categories because it appeared to be the point at which discounting began to level off. Besides price promotions, there are other activities to attract consumers, including book reviews, bestseller lists in print media, direct advertising to consumers, as well as publicity events.7 We analyze the UK book industry around the time when the merger between two of the major book retailers, Waterstone’s and Ottakar’s, took place. Table 1 reports the national market shares in 2005, the year the merger was announced. The main trends, up to 2005, were a sharp growth in the market share of supermarkets and online retailers (both increased by 4% between 2001 and 2005), and a decrease in the share of non- Internet distance sellers (principally book clubs).
[Insert Table 1 here]
At the time of the merger the combined share of the merging parties was 24%. The combined market share of the four largest retailers (i.e. WHSmith, Waterstone’s, Ottakar’s and Borders) was 45%; 55% if only deep-range books are considered.
2.1 The Merger
In August 2005, Waterstone’s and Ottakar’s, announced their intention to merge. At the time of the merger announcement, Waterstone’s was controlled by the HMV Group, a global entertainment retail chain. Waterstone’s had 190 stores in the UK, each with a selection of titles, generally, between 30,000 and 40,000. Ottakar’s was established in 1987 with the aim of creating a chain of bookshops in market towns throughout the UK. It grew both organically and through acquisitions, reaching 141 stores by December 31, 2005. Its stores averaged between 20,000 and 30,000 titles. The merger had mainly a defensive rationale. According to the parties, brick-and- mortar retailers “were caught in a ‘pincer movement’ between, on the one hand, the supermarkets offering a limited range but substantial discounts, both on best-sellers and on an increasing number of deep-range titles, and, on the other hand, the Internet retailers offering a very extensive range. The proposed merger would enable Waterstone’s to compete more effectively with Borders, WH Smith, the Internet retailers and the
7 See Sorensen (2007) for an assessment of the impact of bestseller lists on book sales.
7 supermarkets, as well as other specialist retailers.” (“HMV Group plc and Ottakar’s plc - Proposed acquisition of Ottakar’s plc by HMV Group plc through Waterstone’s Booksellers Ltd”, 12th of May 2006; p. 13). Moreover, the parties stressed the role of potential efficiency gains from introducing Waterstone’s more efficient stock management system into Ottakar’s operations, and from economies of scale whose nature is confidential in the decision (ibid. p. 13). Therefore, as also suggested by the Office of Fair Trading, the merger might be seen as an attempt by “low productivity incumbents … to increase their productivity in response to new competitive pressures” (Office of Fair Trading, 2008, p. 9.) coming from the entry and the growth of supermarkets and online retailers. In the words of HMV chief executive Alan Giles: “A combined Waterstone’s and Ottakar’s business will create an exciting, quality bookseller, able to respond better to the increasingly competitive pressures of the retail market.”8 The CC concluded that the proposed merger was not expected to result in a substantial lessening of competition in the market for the retail sale of new books (best- sellers or deep-range titles) at a local, regional or national level. As a result, the CC cleared the merger unconditionally on May 12, 2006.
3. Data and Sample Selection
To perform our empirical analysis, we built two different datasets, one at the store level and one at the national level. We acquired from Nielsen data on the volumes and values of sales for a sample of 200 book titles sold by 60 of Waterstone’s and Ottakar’s stores. For each selected store, Nielsen provided us with weekly figures of volumes and values, as well as data on several book-specific characteristics, from the first week of January 2004 through the last week of December 2007.9 Moreover, we obtained data on nationally aggregated volumes and values for the same sample of titles over the same time-span, separately for the merging parties and for competitors as a whole.
8 Citation from Seawright, S. (31 May 2006), The Daily Telegraph, http://www.telegraph.co.uk/finance/2940039/Ottakars-falls-to-Waterstone.html, visited May 30, 2014. 9 Throughout the analysis we aggregate data at the monthly level.
8 Because of the peculiarities of the market for books discussed above, we look at the effect of the merger on the prices of a selected sample of titles that varies year by year instead of comparing the same set of titles before and after the merger. Thus, the price of different books is modeled by means of a hedonic approach as a function of the products’ characteristics (e.g. Pakes, 2003).10 This enables us to more correctly identify the effect of a policy change –i.e. the merger decision– on the price level, since any price difference due to the changes in the products’ characteristics is accounted for in the regression. To complete our data, we also gathered information from public sources on socio- economic characteristics (such as population, GDP, Internet penetration, etc.) –both at the local and national levels– that we used for the selection of the 60 stores covered in our analysis as well as additional control variables in the econometric analysis. Sections 3.1 and 3.2 describe how we selected the sample of stores and titles. A description of control variables is provided in section 3.3. A more in-depth analysis in provided in Appendix 1.
3.1. The Choice of the Stores
The DiD analysis at the local level requires the characterization of the treatment and control groups. We define the treatment group as the Waterstone’s and Ottakar’s stores in overlap areas, i.e. local areas where both chains were present before the merger, and the control group as the Waterstone’s and Ottakar’s stores in non-overlap areas. In defining local areas, we take the geographic market definition from the CC’s decision, which focuses the assessment of competition on “nearby locations.” Therefore our overlap areas are those in which Waterstone’s and Ottakar’s were competing head-to-head before the merger. With the exception of one area, stores are located within 0.2 miles from each other (see Appendix 1 for further details). For 2004-2007, the Nielsen Bookscan panel contains a total of 359 Waterstone’s and Ottakar’s stores located in 203 different areas (at the local authority level), of which 33 are the overlap areas identified by the CC and 170 are non-overlap areas. Since it was
10 A similar problem is faced by Ashenfelter et al. (2013), who analyze the price effects of a merger between two appliance manufacturers in the US. Dealing with products with short lifetimes, they also use a model with product characteristics to account for product quality. Unlike them, however, we explicitly build a post-merger sample of titles that reflects the distribution of the observable characteristics in the entire population of titles. We employ a similar methodology in the assessment of the effects of another merger between two videogame retailers in the UK (Aguzzoni et al., 2014).
9 unfeasible to obtain data on all areas due to budget constraints, we built a sample of 60 stores in which the number of Waterstone’s and Ottakar’s outlets is equally split between overlap and non-overlap areas (i.e. 30 and 30). For the overlap areas, we draw the 30 stores from 20 different areas. We select one store for each chain in 10 overlap areas. Then, to increase the coverage of overlap areas, we draw five additional Waterstone’s stores from other five overlap areas and the last five Ottakar’s stores from a different set of five overlap areas. Then, for the control group, we select 30 non-overlap areas: 15 in which we observe only Waterstone’s stores and 15 with only Ottakar’s stores. To summarize, the 60 stores in our sample are selected from a total of 50 areas of which 20 are overlap areas and 30 non-overlap areas. Once the number of areas and the type of stores required was decided, we chose a method to select the specific areas to be included in our sample. The key challenge is to choose areas for the control group that closely resemble the ones chosen for the treatment group in terms of demand and supply conditions such that any post-merger difference between the two groups can be attributed to the merger. Hence, we use a matching methodology that allows us to select two groups of areas with homogeneous observable characteristics and for which we assume that the non-observable characteristics are similarly distributed. In appendix 1 we present a detailed description of the matching methodology.
3.2. The Choice of the Titles
The CC recognized that the merger may have had a different impact on bestsellers and on deep-range titles. Indeed, while bestselling titles faced strong and growing competition, in particular from supermarkets, non-specialist stores, and Internet retailers, deep-range titles appeared less affected by these competitive constraints because supermarkets and non-specialists stores stocked only a small number of such titles. Therefore, we want to account for these possible differences in our analysis. Among books in the top 5,000 in terms of annual sales, there are sizable differences with respect to the volumes sold. For instance, in 2007 the volume of the 200 most sold titles represented around 33% of the total sales of bestsellers, with an average of more than 214,000 copies sold. On the other hand, the 200 books ranked from 4,801 through 5,000 represent just 1% of the total sales, with an average of 7,100 copies sold
10 per title. Hence, we suspect that the retailers’ pricing policy may significantly differ across these 5,000 titles. First, we identify a set of titles that may be characterized by a particular pricing policy, which we call “evergreen.” We include in this category those titles that were ranked among the 5,000 most sold books across the entire period under analysis. In principle, they are subject to unique competitive pressures because evergreens are sold across a variety of retail channels that otherwise devote limited space to books–i.e. non- specialist retailers and supermarkets– as they can guarantee a quite stable flow of sales. Of the remaining books, we define “top-sellers” those titles that were ranked among the 200 most sold books in any given calendar year. Indeed, supermarkets –which are the strongest price competitors of the specialist retailers– tend to concentrate their offers for those books that many consumers are interested in purchasing and have a high position in the sale rankings. Therefore, the pricing strategy of the merging parties for these titles might be different from the one adopted for other frequently sold books. This leaves us with the “bestseller” category, which includes all top 5,000 titles that are not already defined to be “evergreen” or “top-seller.” The last category, “deep-range” identifies all books with a total sales ranking higher than 5001. Other than those books that have earned the status of “evergreen” in our sample, it is important to clarify that titles can change category from one year to the next, varying between bestseller, top-seller, and deep-range. To account for this fact, we selected a sample of titles that are representative of the different categories for each of the years under examination. This results in an unbalanced panel, where some titles are observed for the entire period, while others are only observed for some years. In order to assess any potential effect of the merger that may affect only books with a specific characteristic and to offer a reasonable representation of the universe of books sold, we include in our sample titles with different types of binding and genre.11 We complete our dataset by collecting information for each title in the sample on a
11 The type of binding represents a way of discriminating among consumers with heterogeneous valuation (Clerides, 2002).
11 number of title-specific characteristics such as the date of publication, the number of pages, whether it is part of a series and whether it contains figures. Taking into consideration all the above criteria, we asked Nielsen to randomly select 200 titles subject to the following conditions. At least 20 titles had to be evergreen; for each year at least 10 books had to be top sellers; around 50% of our sample should consist of deep range titles (to reflect CC’s concerns over this type of titles); each year some newly published books should be introduced. In addition to these constraints, we also made sure that the distribution of the main characteristics for each title category in our sample was similar to the distribution of characteristics in the population of all books sold in the UK. Table 2 reports the number of titles included in each category for each year. The annual size of the sample increases over time as titles published after 2004 are progressively added to the sample. Appendix 2 lists all 200 titles on which we perform our econometric analyses, along with some additional descriptive statistics.
[insert Table 2 here]
3.3. The Control Variables
In addition to the title-specific control variables described in the previous subsection, we also built a large dataset of variables to control for the areas’ demand and supply conditions. With respect to the demand side, we collected information on (i) population; (ii) population density; (iii) average sales of books in volumes; (vi) gross value added (GVA); (v) number of universities; (vi) level of education; and (vii) the diffusion of Internet sales (which we proxy through the level of Internet penetration). The first four variables are mainly aimed at controlling for the dimension of the market, while the latter should provide an indication on the local population’s propensity to buy books –i.e. the demand conditions. With regard to the supply side, we gathered data on (i) potential cost shifters such as the cost of paper and average house prices as a proxy of the cost incurred for opening a store; and (ii) a measure of the intensity of competition –i.e. the number of different retailers operating in a given area and their entry and exit rates. Except for the cost of paper, which is an internationally traded commodity, as well as for area-specific variables (such as those related to the intensity of competition), all other variables were
12 collected both at the local and the national level. All variables are described more in- depth in table 3.
[insert Table 3 here]
4. Empirical Analysis
The analysis we undertake aims at evaluating the effects of the Waterstone’s/Ottakar’s merger on the retail market for books in the UK. Our main econometric analysis focuses on the effects of the merger on the price dimension, although we also discuss other non- price effects in Section 4.4 below. The dependent variable of interest is the discount applied to the RRP, because this is the price-related variable retailers compete on. Nonetheless, for the sake of simplicity, we often refer to prices and price competition in the text. Table 4 provides descriptive statistics for the set of variables used in the local analysis. [insert Table 4 here]
Our main econometric analysis examines the impact of the merger at the local level. It compares prices in areas that should have been more affected by the merger (overlap areas) with prices in areas that should have been less affected by the merger (non-overlap areas) using the DiD methodology. Of course, this analysis is only meaningful in the presence of some price variability at the local level. Hence, as a preliminary step, we illustrate the evidence supporting the existence of price variability across local markets.
4.1 Local Price Variation
The existence of local competition in the UK retail book market was undisputed and it was central to the entire CC investigation. Both the CC and the parties undertook distinct analyses to explore the geographic extent of local competition, but none thoroughly examined the issue of local pricing. However, in the text of the CC’s decision some evidence suggests the existence of local variation in price promotions, particularly in bundle sales, such as “3 for 2”, that are common across retailers. First, as far as Waterstone’s is concerned, the CC found that “different multibuy arrangements can be set
13 up for different areas…” (“HMV Group plc and Ottakar’s plc - Proposed acquisition of Ottakar’s plc by HMV Group plc through Waterstone’s Booksellers Ltd”, 12th of May 2006; Appendix D, p. D2). Second, with regards to Ottakar’s, the CC found that: “There are centrally determined local variations in the 3 for 2 promotions. There may be a different 3 for 2 for large stores and small stores … or for London or Scotland.” (ibid. Appendix D, p. D4). Therefore, even though local managers might in fact exert their discretion mainly in the choice of the range of titles to be stocked in the store and in non- price promotions –such as giving books more prominent display space– there seems to be a certain amount of local variation in price promotions. Moreover according to the CC’s decision, local managers also have discretion in relation to the sale of dead stock and genre promotions. To further investigate the extent to which prices varied across stores, we compute the distribution of the coefficient of variation of the discounts granted by Waterstone’s and Ottakar’s across the 60 stores in our sample for each title and for each month. For each title and for each month in our sample, the coefficient of variation was calculated as the ratio between the standard deviation and the mean of the average percentage discount across the 60 stores. The coefficients of variation provide a measure of price dispersion across stores that is not affected by differences in the average prices of titles. Figure 1 presents the distribution of the coefficient of variation of the discount across stores for the four categories of titles. Figure 2 shows the distribution of the coefficient of variation of the discount across stores for each of the four categories of titles pre- and post-merger. [insert Figure 1 here] [insert Figure 2 here]
The analysis of the coefficients of variation confirms the existence of some discount variability across stores, both before and after the merger. The only category showing a distribution that is slightly different from the others is the top-sellers’, for which the discounts are on average higher since they are subject to tougher competition than the other categories because new releases are relatively more price-promoted on a national basis.
14 One particular challenge we face when looking at prices is related to the way bundled sales are recorded in our dataset. Titles included in the bundles are sold at a greater discount than the one granted to stand-alone sales. This implies that the more titles a store sells through bundles, the lower is its average selling price. Some price dispersion across stores may then be caused by different rates of bundle take-up, rather than by different pricing strategies. As Nielsen does not collect information on whether a book is sold as stand-alone or as part of a bundle, we cannot fully control for the variability in the data that is due to different rates of bundle take-up between stores. Nonetheless, we look at more disaggregated measures of price dispersion (see Appendix 3 for further details) and we consistently find evidence of price variability across stores, which suggests that prices competition, at least to some extent, took place at the local level.
4.2 Average Treatment Effect
In Figure 3, we plot the monthly average discounts in the overlap and non-overlap locations for the different categories of titles. If the merger increased the price –i.e. reduced the discount– we would expect to observe an increase in the vertical distance between overlap and non-overlap lines after the merger. Yet apparently, discounts in overlap areas tend to follow broadly the same pattern as those in non-overlap areas and do not seem to be systematically lower post-merger.12 We further formally test the hypothesis that there is no differential trend between overlap and non-overlap areas prior to the merger, which is crucial for the validity of our DiD estimation (see appendix 4 for details). [insert Figure 3 here]
The identifying assumption at the basis of our DiD analysis is that, if local managers were, at least to some extent, free to set prices at the store level and the merger had anticompetitive effects, these effects should have been larger in the overlap areas,
12 The more pronounced volatility for top-sellers is partially due the fact that our sample for these titles is small, which implies that a change in the discount that is applied only to a few titles may significantly affect the average. In particular, in the first four months of 2007 we have only one top-selling title in our sample, which was sold at a very high discount. Therefore, when running the econometric exercise on top- selling titles, we exclude the data for this period.
15 because of the reduction in local competition brought about by the merger. We therefore estimate the following regression:
disc ist D E post t O overlap s G post t u overlap s J X i P Z st K t Q is H ist (1)
where discist is the discount on the recommended retail price on title i granted in store s at
time t, postt is a dummy equal to 1 for the titles observed in the post-merger period and 0
before, overlaps is a dummy equal to 1 for the titles sold in overlapping stores and 0
otherwise, Xi is a set of title-specific control variables, Zst is a set of variables aimed at
controlling for changes across time in local market features, Șt is a time trend, Ȟis is a title- 13 store fixed effect and İist is the error term.
Our key variable is the interaction between postt and overlaps, whose coefficient (G) measures the price change in overlap locations relative to the price change in non- overlap areas: the average treatment effect (ATE). This coefficient quantifies the additional variation experienced by the prices in the overlap areas with respect to the average price change in the non-overlap areas. The post coefficient (ȕ) measures any price change (between the pre-merger and the post-merger period) common to all locations, while the coefficient Ȝ, related to the overlap regressor, accounts for any idiosyncratic differences between overlap and non-overlap areas that are not related to the merger. In appendix 4, we show the results of an alternative specification that includes time period dummies instead of a post-merger dummy, which are not qualitatively different from those of our chosen specification. That specification also allows us to verify that there is no differential trend between treatment and control groups prior to the merger, which is crucial for the estimation of our treatment effect.In order to do this, we plot the coefficients on the interaction terms against the corresponding calendar date and then run a formal test of the null that there are no differential trends (see Appendix 4 for further details).
13 We cluster the error terms at the title and store level in order to control for autocorrelation in the error process. As a robustness check, we also present results of a specification with standard error clustered at the store level in Appendix 4. The main results are not affected. As a further robustness check we also estimate regressions in which we impose an AR(1) error structure on the model, and the resulting estimates are similar to those obtained by clustering the error terms.
16 We perform a pooled regression on all the titles together in order to quantify the overall effect of the merger. We also run the regressions on each category separately, as the different categories of titles face different competitive conditions and are therefore expected to exhibit dissimilar discount patterns. In our baseline specification, we run a regression with fixed-effects for each title/store, so as to capture all the time-invariant title/store-specific (unobserved and observed) characteristics that may affect their prices. However, the use of title/store-specific fixed-effects implies that the effect of the merger on discounts is solely identified from titles sold both before and after the merger, because the interaction variable is perfectly collinear with the title/store fixed-effects for those books whose prices were observed only before or after the merger. This may affect the estimates because it reduces the size of the sample of titles on which the effects of the merger are actually measured and it does not allow us to capture any change in the prices of the titles published after the merger. In our sample this problem arises from two different sources: (i) some of the titles included were published after the merger; and (ii) some of the titles included changed category over the period examined, which is clearly only relevant when we run the category-specific regressions. The titles that belong to (i) are mainly top-selling titles and, in very few cases, bestsellers and deep-range books, while for the titles which belong to (ii) the problem spans across all the categories, since titles frequently move up or down the rankings from one year to the next. When we pool all titles together, the fixed-effects problem is not a big concern since the potential distortion comes only from titles that are published after the merger. These titles represent only a small fraction of the sample (19 out of 200 titles). Hence, in the pooled regressions we opt for a specification with fixed- effects at title/store level. When instead we run the regressions for each category of titles separately, the fact that a large proportion of titles change category over time, and in particular before and after the merger, poses a serious problem to the fixed-effect estimation. To overcome this problem, instead of the title/store-specific fixed-effects, we include a large set of observable characteristics (both title-specific and store-specific) that
17 may affect a title’s price, using a random-effect specification within a hedonic pricing approach.14 A further methodological issue we address is related to the selection of the window of data surrounding the merger to be excluded from the analysis, since we do not know when the merging parties started operating as a single entity. We consider two possible windows (6 and 12 months) around the date of merger clearance. We find that the results are essentially unaffected by the size of the window. Therefore, we only report results based on the window dropping the fewest observations, i.e. the one excluding only the 6 months around the merger date.In Appendix 4, we also show a robustness check where we include data within the merger window and allow the treatment effect to vary within this time period. Results are in line to those of our main specification. Table 5 reports the results of the DiD regression (1). In the first column, we use the full sample of all titles with fixed-effects for each combination of stores and titles. The coefficient for post×overlap is not significantly different from zero. Therefore, the merger does not seem to have had a different impact in overlap and non-overlap areas. The 95% confidence interval shows that the range of effects is between -0.70% and +0.12%. Therefore, the range of effects that can be ruled out seems to be quite relevant in the context of an average discount of 10%. Columns 2 to 5 report the results of the regressions run on best-selling titles, deep-range titles, evergreen titles, and top-selling titles, respectively. Again, the coefficient estimates for post×overlap, are never statistically significant (the 95% confidence intervals ranging between -1.07% and 0.94% for bestsellers, -0.57% and 0.28% for deep-range, -0.60% and 0.82% for evergreens, and -1.57% and 1.28% for top-sellers.), which seems to confirm that the merger did not adversely affect the discounts applied by the merging parties in the overlap areas, on average. [insert Table 5 here]
14 The random-effect specification has a potential drawback as the estimator can be biased if there are unobservable characteristics that systematically changed after the merger occurred. Following the approach of Ashenfelter et al. (2013), we run our baseline regression on the sample of titles that remained in the same category before and after the merger using both a fixed-effect specification and a random-effects specification with title and store characteristics. The results of this comparison are presented in Appendix 4. We find that the estimate of the treatment effect is similar under both specifications, thereby suggesting that unobservable product characteristics do not significantly bias the estimator of the random-effect regression.
18 The coefficient of the post variable suggests that there is a statistically significant upward shift of the time trend after the merger (+2.4%) for deep-range titles, and a downward shift (-3.5%) for top-selling titles. The coefficients of the variables that measure the intensity of local competition, such as the number of competing retailers that are active in the area and the number of competitors that entered/exited the market in the previous three months, suggest that local competition variables do play a role in determining prices (see Appendix 4 for further details). We do not extensively report results on all other controls, which mostly conform expectations. In particular, we observe some common and statistically significant effects relative to the title characteristics. First, paperback titles are associated with lower discounts compared to hardcover titles (except for the top-selling titles). Second, the discounts appear to be lower (except for deep-range books) as the time elapsed from the publication increases. Third, the publication of a new title is on average accompanied by promotional discounts (again except for the deep-range category). Fourth, when a book contains figures the discount is on average lower. Fifth, titles that are part of series are usually sold at a higher discount. Finally, around Christmas the discounts tend to be higher (except for deep- range titles). It is also interesting to notice that the positive and significant coefficient for the Waterstone’s dummy suggests that Waterstone’s stores applied on average higher discounts than Ottakar’s. The remaining control variables included in the model are, instead, mostly not significant and, even when they are, the sign of the coefficients differs across categories. In conclusion, the merger does not seem to have on average adversely affected prices in the overlapping areas where it could have been expected to generate the strongest effects due to the increase in the level of market concentration.
4.3 Heterogeneous Treatment Effect
The fact that we do not observe any significant average result might just mean that our simple framework fails to model some important underlying heterogeneity in the merged entity’s conduct post-merger. Hence, we perform heterogeneous treatment effects estimations to more precisely investigate the impact of the merger on prices. In particular, we assess whether the effect of the merger differs between overlap and non-overlap areas
19 along three dimensions: book-specific, firm-specific, and market-specific. There are good reasons to expect possible heterogeneous effects of the merger along these dimensions. The merger might for instance only affect the merging firms’ pricing strategy for the new products released after the merger as compared to old products; it can affect the extent of price differentiation between the merging firms if they internalize the externalities they exert on each other’s products; and finally it might have differential effects depending on the competitive conditions of local markets. Table 6 reports the estimated coefficients for several heterogeneous treatment effects. In the first set of regressions our key variable post×overlap is interacted with firm-specific dummy variables. The results of these regressions suggest that, after the merger, Waterstone’s stores reduced the discounts by 1.5% on average in overlap areas, while Ottakar’s stores increased them by 0.9%. Together with the finding of our baseline specification in Table 5, according to which, before rebranding, Waterstone’s stores applied on average a discount 1.6% higher than Ottakar’s, this implies that there has been a convergence of the prices between the two merged chains in overlap areas after the merger.
This price convergence might have important policy implications as it has long been recognized that uniform prices can have a mixed effect on consumers’ surplus (e.g. Hausman and Mackie-Mason, 1988): some customers might be better off while others are worse off. Yet, to draw clear welfare conclusions about this change in pricing strategy would require an analysis of sales volumes for all book titles, which is unfortunately infeasible with the data at hand.
[insert Table 6 here]
We further consider whether the effect of the merger mainly materializes in the overlap areas where the merging parties closed a store after the merger, although there were very few occurrences of store closures by the merged entity after the merger (see Appendix 1 for a list of locations where there was a store closure). Yet, the coefficient post×overlap×closed, which captures the effect of the merger in those areas, is not significantly different from zero, thereby indicating that not even in those locations where
20 there was a reduction in the number of stores due to the exit of one of the merging parties can we observe a systematic difference in the prices after the merger.15 This finding is in line with the CC’s expectations at the time of the decision. The reasons why the CC considered that store closures would not hinder competition are mainly related to the existence of Internet retailers and other brick-and-mortar stores, as well as the fact that any possible reduction in range due to store closure would probably be offset by the introduction of the improved stock management system at Ottakar’s, which would lead to an increase in range of 7-10%. The second set of regressions deals with title-specific heterogeneity to explore whether the merger influences products existing prior to the merger differently from products introduced after the merger. In our sample, we observe (see Section 4.2) that retailers tend to grant higher discounts on titles that have been just published –i.e. in the two months following publication. The merger may have then had an impact on the timing of discounts. To capture this effect, we partition the treatment effect into three components: (i) the component capturing changes in prices for those titles released before the merger (post×overlap×released_pre); (ii) the component capturing changes in prices for those titles released after the merger but only considering the first two months after release (post×overlap×released_post×just_pub); and (iii) the component that is identified by the changes in prices for those titles released after the merger, but only considering the period following the second month after release. Since all the titles published post-merger essentially belong to the top-sellers category, we only show estimates of heterogeneous treatment effects for this category. The positive and significant coefficient of post×overlap×released_post×just_pub suggests that stores in overlap areas seem to price newly released books more aggressively after the merger. Instead, the negative and significant coefficient of post×overlap×released_post×non_just_pub indicates that stores in overlap areas set significantly higher prices for titles published after the merger two months after the release date. This result seems to suggest that the merger might have also affected the
15 The 95% confidence interval for specification (1) ranges between -1.28% and 0.55%. We do not report all other confidence intervals due to space limitation.
21 firms’ incentives in relation to the timing and duration of discounts. Yet, this result should be taken with caution given the limited number of titles released post-merger in our sample. Nonetheless we think that the mergers’ impact on the timing of discounts/prices in short-life product markets may be a fruitful line for future research. The third set of heterogeneous treatment effect regressions deal with a market- specific heterogeneity. Similarly to Hosken et al. (2012), we look at whether the effect of the merger on prices differs according to the intensity of competition as measured by the number of competitions present in the market before the merger.16 In particular, we define the variables post×overlap×high_comp, post×overlap×medium_comp and post×overlap×low_comp, which represent the effect of the merger in overlap areas with high, medium, and low number of competitors respectively, based on the 33rd and 66th percentile of the distribution of the number of competitors in all the areas. Moreover, we investigate whether the merger had a differential impact depending on the dynamics of entry and exit. More precisely, we define the following variables: post×overlap×high_entry/ post×overlap×high_exit represents the effect of the merger in the areas where the entry/exit of 3 or more competitors occurred after the merger; post×overlap×medium_entry/ post×overlap×medium_exit represents the effect in the areas where there was entry or exit of 1-2 competitors after the merger. The expectation is that entry can mitigate the potential anti-competitive effects of the merger. These latter interactions should be also interpreted cautiously as they might be potentially affected by an endogeneity problem (i.e. entry and exit may be triggered by the pricing conduct of the merged entity following the transaction). Nevertheless, the coefficient estimates for these two sets of interaction variables are mostly not significant, indicating that market-specific characteristics that should capture differences in the areas’ competitive conditions do not seem to play a significant role in explaining the effect of the merger at local level.
16 Unfortunately, information on market shares in local areas is not available in our data. Hence, we must rely on the simple count of the number of competitors instead of a more precise measure of market concentration as used by Hosken et al. (2012) to analyze the heterogeneous effect of mergers in the US grocery retailing markets.
22 4.4 Analysis of Sales Volume
Although our analysis focuses on prices, we reckon that non-price variables, such as product variety of service quality, may play an important role. Indeed, Inderst and Shaffer (2007) show in a theoretical model that retail mergers may reduce product variety. Unfortunately, Nielsen does not collect information on variables such as service quality, store amenities and product differentiation across shops, nor it was possible to obtain such information form the merging parties. We could not therefore directly assess the impact of the merger on these variables. However, as suggested by one referee, if the merger reduced quality or variety we may observe a reduction in the volume of the merging parties’ sales. We have therefore analysed the impact of the merger on monthly average volumes at the title level. We used the same specification and the same set of regressors that we used in the price equation (equation 1). As reported in table 7 and consistently with the findings on prices, the merger did not seem to have any effect on volumes when comparing overlap and non-overlap areas. These findings provide indirect evidence of the fact that the merger did not negatively affect either quality of service or product variety.
[insert Table 7 here]
5. National Analysis
So far our analysis has focused on the effect of the merger on the local competition. However to the extent that the parties mainly set nationally uniform prices, the transaction could also have had an aggregate effect on national prices, which would not be captured by our local analysis. We perform a complementary DiD analysis to investigate the merger’s effect at the national level. A major issue in implementing this alternative analysis is the identification of a suitable counterfactual. To improve the robustness of our results, we employ two different control groups: (i) the same titles sold by the competitors of Waterstone’s and Ottakar’s; and (ii) the top-selling titles sold by the merging parties. We describe the rationale behind both methodologies in turn, while descriptive statistics and additional analyses are presented in Appendix 5.
23 5.1 Competitors’ Titles as Control Group
As a first control group for our DiD analysis with nationally aggregate prices, we use the prices charged by the rival firms. Using competitors’ products as a control group is a common practice in this literature (see, among others, McCabe, 2002). Indeed exogenous supply or demand shocks affecting the whole industry should be expected to hit in a similar way the prices of the merging parties and those of their competitors. However, if firms compete on prices, the discounts applied by all retailers in the market are likely to be correlated and, thus, the merger may affect not just the discounts granted by the parties, but also those granted by their competitors. Yet, according to the standard theoretical merger model by Deneckere and Davidson (1985) where firms compete in prices and goods are differentiated, any price change by the merging parties post-merger should be followed by a price change in the same direction by their rivals, but of a lesser magnitude. Comparing the change in prices of the merging parties to that of the competitors may therefore provide a useful indication as to whether the merger produced negative price effects. In this case the general estimation equation is:
discijt D E postt O mergedj G postt umergedj J Xi P Zt Kt Qi Hijt (2)
where discijt is the discount on the recommended retail price on title i granted by retailer j at time t, postt is a dummy equal to 1 for the titles observed in the post-merger period and
0 before at retailer j, mergedj is a dummy equal to 1 for the titles sold by the merging parties; 0 otherwise; and measures the time-invariant difference between the merging parties and their competitors. Xi is a set of title-specific control variables and Zt is a set of variables aimed at controlling for changes across time in the demand and supply 17 18 conditions at the national level, Șt is a time trend, Ȟi is a title fixed-effect, and İist is the error term. The key variable of interest is the interaction between postt and mergedj,
17 As we discuss in Appendix 5, we impose two distinct trends, one for the merging parties and one for the competitors, so as to isolate the effect of diverging trends of discounts from that of the merger, which seem to be due to structural changes in the supply-side of the market. 18 Similarly to the local analysis, in the regressions by category we replace the fixed-effect with product characteristics. A comparison of the estimated price effect using these two different specifications on the sample of titles that remain in the same category before and after the merger is presented in Appendix 5.
24 whose coefficient (į) measures the price change of the merging parties relative to the price change of competitors attributable to the merger.
[insert Table 8 here]
Table 8 presents the results of the estimation of equation (2). In the pooled regression (column 1), the coefficient’s estimate for post×merged is not significant suggesting that the merger does not appear to have had any differential impact on discounts, either between the merging parties or their competitors. Consistent with the results of the pooled regression, the analysis at the category level (columns (2) to (5)) shows that the coefficient’s estimate for post×merged is never significant. Overall, this analysis suggests that the merger did not affect national prices.
5.2 Top-selling Titles as Control Group
To overcome the issues that arise when using competitors’ titles as a control group, we also perform an additional DiD estimation in which we use the top-selling titles as a control group. The top-sellers appear to be the category where the merging parties face the fiercest competition, as these titles are sold by all types of retailers and, in particular, by supermarkets, which have the most aggressive pricing policy. Therefore, the merger could be expected to have the most limited effect on the prices of these titles. The estimating equation is then as follows: discikt D E postt O non _ topk G postt u non _ topk J X i P Zt Kt H ikt (3)
where discikt is the discount on the recommended retail price on title i in category k at
time t, Į is a constant, postt is a dummy equal to 1 for the titles observed in the post-
merger period and 0 before for category k, non_topk is a dummy equal to 1 for the titles in category k other than top-sellers and captures the systematic difference –i.e. not related to the merger– between top-selling titles and other categories of titles, and the coefficient į captures the effect of merger on these latter book categories –i.e. evergreen, bestseller, and deep-range books. This equation is run only on the merging parties’ prices. The results of the DiD analysis with the top-selling titles as the control group equation (3) are presented in table 9. Columns (1), (2), and (3) report the estimates of the
25 regressions with bestsellers, deep-range books and evergreen books as the treatment group respectively. The coefficient estimate for post×non_top that measures the effect of the merger relative to the top category, is never significant, thereby indicating that the merger did not differentially affect the discounts applied to bestseller, deep-range and evergreen titles if compared to top-titles. [insert Table 9 here] Overall, even though some caution in the interpretation of the estimates is required because of the limited size of the post-merger sample for top-selling titles and, hence, small variations on the discount on very few titles may artificially increase/decrease the average discount of the category, our results show that the merger did not produce any negative effect on prices at the national level. This is consistent with the outcome of the DiD regression that uses the prices of the competitors as control group. Finally, we also estimate heterogeneous treatment effects at the national level but we do not find any significant effect (see Appendix 5).
6. Conclusions
The ex-post assessment of merger effects is an important and increasingly used tool to inform and guide the decision making of antitrust agencies in prospective merger cases. Despite the large number of mergers in the retailing sector that antitrust authorities have decided upon in recent years, there is lack of empirical work estimating the effects of consummated mergers in these industries. Our paper tries to fill this gap by econometrically analyzing the price effects of the merger between two major bookstore chains in the UK –Waterstone’s and Ottakar’s. A peculiar characteristic of mergers in retail industries is the fact that they can exert their influence at different geographical levels since retail chains may set their pricing policies either at the national or local level due to the dispersed distribution of buyers and sellers. Unlike previous retrospective merger studies in other sectors, our empirical framework makes use of this feature to identify more precisely the effect of the merger. The availability of data at both the local and national level coupled with different identification strategies and, in particular, different counterfactuals, allows us to perform complementary assessments of the effect of the merger, which give a broader picture of its implications.
26 In our main econometric analysis, we perform a DiD analysis where we compare the price change before and after in areas where both chains are present before the merger (overlap areas) to the price change before and after the merger in non-overlap areas where only one of the merging parties is present. Our results show that the merged parties, in aggregate terms, did not significantly increase the prices in those overlapping locations where they might have been expected to do so. We also assess whether the merger lead to an increase in price at an aggregate/national level. The results of two different DiD analysis – one with competitors as control group, and one with top-selling titles as control group– show no significant effect of the merger on national prices. We further exploit the richness of our data to empirically identify a differential response to the concentration by the merging parties. After the merger, Ottakar’s –the perceived premium chain– significantly decreased prices while Waterstone’s significantly increased them in the overlap areas. We also find that the timing of the discounts for the top-selling titles released after the merger in the overlap areas seem to differ from those released in the non-overlap areas even though this effect is only identified by few observations. Our findings are consistent with the conclusions that the CC drew after its investigation of the merger. As several competing retailers, especially supermarkets and online outlets, were imposing an increasingly relevant competitive constraint on the merged entity, the CC considered it unlikely that the parties would be able to increase prices following the merger. Arguably, the merger might have affected competition along other dimensions. Specifically, it might have led to a reduction in the range of titles on offer and/or deterioration in some quality aspects of retailers’ offer. Although these aspects do not seem to be crucial in the Waterstone’s/Ottakar’s merger because the CC deemed it unlikely that the merged entity would reduce the range or service quality, they might be relevant in other retail industries and therefore should be considered when evaluating ex-post the impact of a merger.
27 References Aguzzoni L., E. Argentesi, P. Buccirossi, L. Ciari, T. Duso, M. Tognoni, and C. Vitale, 2011, “The Ex-post Evaluation of Two Merger Decisions”, a report prepared for the UK Competition Commission. Aguzzoni L., E. Argentesi, P. Buccirossi, L. Ciari, T. Duso, M. Tognoni, and C. Vitale, 2014, “A Retrospective Merger Analysis in the U.K. Videogame Market”, Journal of Competition Law and Economics, forthcoming. Ashenfelter O., and D. Hosken, 2011, “The Effect of Mergers on Consumer Prices: Evidence on Five Mergers on the Enforcement Margin”, Journal of Law and Economics, 53 (3): 417-466. Ashenfelter O., D. Hosken and M. Weinberg, 2009, “Generating Evidence to Guide Merger Enforcement”, Competition Policy International, Volume 5, No. 1. Ashenfelter O., D. Hosken and M. Weinberg, 2013, “The Price Effects of a Large Merger of Manufacturers: A Case Study of Maytag-Whirlpool”, American Economic Journal: Economic Policy, 5 (1): 239-61. Beck J., 2004, “Fixed, Focal, Fair? Book Prices Under Optional Resale Price Maintenance”, WZB Working Paper, SP II 2004 – 15. Beck J., 2007, “The Sales effect of Word of Mouth: A Model for Creative Goods and Estimates for Novels”, Journal of Cultural Economics, 3(1): 5-23. Björnerstedt J. and F. Verboven, 2012, “Does Merger Simulation Work? A "Natural Experiment" in the Swedish Analgesics Market”, CEPR Discussion Paper No. DP9027. Canoy et al., 2006, “The Economics of Books”, in Handbook of the Economics of Art and Culture, by Ginsburgh V. and C. D. Throsby (eds.), vol. 1, pp. 721-761. Chandra A. and A. Collard-Wexler, 2009, “Mergers in Two-Sided Markets: An Application to the Canadian Newspaper Industry”, Journal of Economics and Management Strategy, 18(4), 1045-1070. Choné P. and L. Linnemer, 2012, “A Treatment Effect Method for Merger Analysis with an Application to Parking Prices in Paris”, Journal of Industrial Economics, 60 (4): 631-656. Clerides S., 2002, “Book Value: Intertemporal Pricing and Quality Discrimination in the US Market for Books”, International Journal of Industrial Organization, 20 (10): 1385–1408. Competition Commission and Office of Fair Trading, 2011, “Commentary on Retail Mergers”, available at www.oft.gov.uk/shared_oft/mergers/oft1305-ccV1a.pdf. Davies S. W., H. Coles, M. Olczak, C. Pike, and C. Wilson, 2004, “The benefits from competition: some illustrative UK cases,” DTI Economics Paper No. 9.
28 Davis P., 2006, “Spatial competition in retail markets: movie theaters”, The RAND Journal of Economics 37 (4): 964-982. Deneckere R. and C. Davidson, 1985, “Incentives to form coalitions with Bertrand competition”, Rand Journal of Economics, 16 (4): 473-486. Duso T., 2012, “A Decade of Ex-post Merger Policy Evaluations: A Progress Report,” in Dan Sjöblom Ed., More Pros and Cons of Merger Control, Stockholm: Swedish Competition Authority, 125-187. Focarelli D. and F. Panetta, 2003, “Are Mergers Beneficial to Consumers? Evidence from the Market for Bank Deposits”, American Economic Review, 93 (4): 1152-1172. Friberg R. and A. Romahn, 2012, “Ex-Post Merger Review and Divestitures”, mimeo, Stockholm School of Economics. Hastings J. S., 2004, “Vertical Relationships and Competition in Retail Gasoline Markets: Empirical Evidence from Contract Changes in Southern California”, American Economic Review, 94 (1): 317-328. Hausman J. and J. Mackie-Mason (1988), “Price Discrimination and Patent Policy,” Rand Journal of Economics, 19: 253-265. Hosken D., L.M. Olson, and L.K. Smith, 2012, “Do Retail Mergers Affect Competition?
Evidence from Grocery Retailing”, Federal Trade Commission, Bureau of Economics, Working Paper No. 31. Hunter G., G.K. Leonard, and S. Olley, 2008, “Merger Retrospective Studies: A Review”, Antitrust, 23 (1): 34–41. Inderst R. and G. Shaffer, 2007, “Retail Mergers, Buyer Power and Product Variety”, Economic Journal, 117 (1): 45–67. Kovacic W., 2009, “Assessing the Quality of Competition Policy: The Case of Horizontal Merger Enforcement,” Competition Policy International, 5, 129–150. McCabe M., 2002, “Journal Pricing and Mergers: A Portfolio Approach”, American Economic Review, 92 (1): 259-269. Nevo A. and M.D. Whinston, 2010, “Taking the Dogma out of Econometrics: Structural Modeling and Credible Inference”, Journal of Economic Perspectives, 24 (2): 69-82. Office of Fair Trading, 2008, “An evaluation of the impact upon productivity of ending resale price maintenance on books,” http://www.oft.gov.uk/shared_oft/economic_research/ oft981.pdf.
29 Pakes A., 2003, “A Reconsideration Of Hedonic Price Indexes With An Application To PC's”, American Economic Review, 93 (5): 1578-1614. Skrainka B. S., 2012, “The Geography of Grocery Demand in the UK: An Evaluation of the 2003 Morrisons-Safeway Merger”, mimeo, University of Chicago. Sorensen A., 2007, “Bestseller Lists and Product Variety”, Journal of Industrial Economics, 55 (4): 715-738. Weinberg M., 2008, “The Price Effects of Horizontal Mergers”, Journal of Competition Law and Economics, 4 (2): 433-447. Weinberg M., 2011, “More Evidence on the Performance of Merger Simulations”, American Economic Review: Papers & Proceedings, 101 (3): 51–55.
30 Figures and Tables
Figure 1: Distribution of the monthly coefficient of variation across stores by category
CV across categories Monthly prices .15 .1 kdensity cv_m .05 0
0 5 10 15 CV
Best Seller Deep range Evergreen Top
Figure 2: Distribution of the monthly coefficient of variation across stores by category before and after the merger
Best sellers Deep Range Monthly prices Monthly prices .2 .2 .15 .15 .1 .1 .05 .05 kdensity cv_m kdensity cv_m kdensity 0 0
0 2 4 6 8 10 0 2 4 6 8 10 CV CV
Evergreen Top Monthly prices Monthly prices .2 .2 .15 .15 .1 .1 .05 .05 kdensity cv_m kdensity cv_m kdensity 0 0
0 2 4 6 8 10 0 2 4 6 8 10 CV CV
Pre-Merger Post-Merger
31 Figure 3: Distribution of the monthly discounts: Overlap vs. non-overlap locations
Best Seller Deep Range Merger window Merger window 40 40 30 30 20 20 10 10 0 0
(% discount over RRP) over discount (% 2004m1 2005m1 2006m1 2007m1 2007m12 RRP) over discount (% 2004m1 2005m1 2006m1 2007m1 2007m12 % discount= (RRP-ASP)/RRP % discount= (RRP-ASP)/RRP
Evergreen Top Merger window Merger window 40 40 30 30 20 20 10 10 0 0 (% discountover RRP) (% discount overRRP) 2004m1 2005m1 2006m1 2007m1 2007m12 2004m1 2005m1 2006m1 2007m1 2007m12 % discount= (RRP-ASP)/RRP % discount= (RRP-ASP)/RRP
Overlap Non overlap
Table 1: National Market Shares of Retailers (based on the value of sales in 2005)
Firm Market shares Waterstone’s 17% Ottakar’s 7% Other specialist bookshops, including Borders and Blackwells 15% Other stores, including WHSmith 19% Supermarkets 8% Internet 8% Book clubs and other distance sellers 15% Other 10% Total 100%
Source: CC’s calculations based on TNS and Nielsen Bookscan data.
32 Table 2: Number of titles by category and year Book Type Bestseller Deep-range Evergreen Top-sellers Total 2004 40 76 20 18 154 2005 65 82 20 12 179 2006 54 106 20 11 191 2007 44 126 20 10 200
33 Table 3: Description of variables
Variable Description Source overlap Dummy=1 if overlap area post Dummy=1 if post merger (after June 2006) discount [RRP – (sales values/sales volumes)]/RRP*100 Nielsen Bookscan month_t Monthly trend closed Dummy=1 if Waterstone’s or Ottakar’s closed a shop in the area season Seasonal dummy for Christmas period waterstone Dummy for Waterstone’s stores (before rebranding) trading_m1 N° of stores for specialist retailers trading_m2 N° of stores for non-specialist retailers trading_m3 N° of supermarkets including other retailers (e.g. DIY chains) that sell books as a part of a wide range of goods entry1 N° of stores for specialist retailers that have entered the market over the last three months entry2 N° of stores for non-specialist retailers that have entered over the last three months entry3 N° of supermarkets including other retailers (e.g. DIY chains) that have entered over the last three months exit1 N° of stores for specialist retailers that have exited over the last three months exit2 N° of stores for non-specialist retailers that have exited over the last three months exit3 N° of supermarkets including other retailers (e.g. DIY chains) that have exited over the last three months Series Dummy=1 for titles which are part of a series Figure Dummy=1 for titles containing figures Pages Number of pages elapsed_year Years elapsed since the publication just_pub Dummy=1 for the first 2 months after publication paperback Dummy=1 for paperback titles avgsales_area Average sale volumes of Waterstone’s and Ottakar’s stores per area (in 2005) woodpulp Cost of paper World Bank internet Internet penetration Internet Access, Households and Individuals, Office of National Statistics house_price House prices http://www.landregistry.gov.uk/ for England and Wales, and www.ros.gov.uk for Scotland population Population UK Office for National Statistics, Neighbourhood Statistics section or Scottish Neighbourhood Statistics when relevant (2001 census) pop_density Population density UK Office for National Statistics, Neighbourhood Statistics section or Scottish Neighbourhood Statistics when relevant (2001 census) universities Number of universities http://www.lovemytown.co.uk/Universities/Universitie sTable1.asp education Average level of education as defined by the UK Office for National Statistics, Neighbourhood National Qualification Framework. These Statistics section or Scottish Neighbourhood Statistics levels range from 1 (secondary education - when relevant (2001 census) GCSE- with marks below or equal to D) to 7 (Doctoral degree). GVA Gross value added. It measures the UK Office for National Statistics, Neighbourhood contribution to the economy of each Statistics section or Scottish Neighbourhood Statistics individual producer, industry or sector. GVA when relevant (2001 census) = GDP–taxes on products +subsidies on products.
34 Table 4: Descriptive statistics of control variables for local analysis
Variable Obs. Mean Std. Dev. overlap 207,690 0.49 0.50 post 207,690 0.39 0.49 discount 207,690 10.58 12.93 month_t 207,690 24.25 13.33 closed 207,690 0.07 0.25 season 207,690 0.19 0.40 waterstone 207,690 0.54 0.50 trading_m1 207,690 0.91 1.45 trading_m2 207,690 9.58 7.19 trading_m3 207,690 7.49 6.15 entry1 204,013 0.02 0.12 entry2 204,013 0.09 0.29 entry3 204,013 0.05 0.22 exit1 204,013 0.01 0.09 exit2 204,013 0.04 0.20 exit3 204,013 0.03 0.18 classD1 207,690 0.31 0.46 classD2 207,690 0.09 0.29 classD3 207,690 0.28 0.45 classD4 207,690 0.31 0.46 series 207,690 0.35 0.48 figure 207,690 0.51 0.50 pages 207,690 301.71 236.79 elapsed_year 207,690 3.34 3.47 just_pub 207,690 0.07 0.25 paperback 207,690 0.90 0.29 avgsales_area 207,690 190,506.50 86,660.03 woodpulp 207,690 5,459.25 512.25 internet 207,690 57.25 6.17 house_price 207,690 195,713.70 65,260.84 population 207,690 182,723.40 102,776.00 pop_density 207,690 14.75 12.78 urban area 207,690 0.86 0.34 universities 207,690 0.85 0.73 education 207,690 20.53 6.02 GVA 207,690 17,953.33 3,979.17
35 Table 5: DiD on Local Prices - Average Treatment Effect
All titles Bestsellers Deep-range Evergreen Top-sellers (1) (2) (3) (4) (5) Overlap 0.369 0.121 0.041 0.503 (0.974) (0.540) (0.122) (1.085) Post 1.361*** 0.987* 2.363*** -0.387 -3.469*** (5.94) (1.881) (9.789) (-1.015) (-3.469) overlap×post -0.290 -0.061 -0.148 0.112 -0.147 (-1.38) (-0.119) (-0.684) (0.310) (-0.202) Constant 18.701*** 10.88*** 14.83*** 5.596*** 11.27*** (5.77) (4.604) (11.23) (2.844) (3.408)
Observations 172,991 37,094 58,098 56,955 20,433 Number of id 11,833 4,544 6,909 2,445 2,930 R-squared 0.061 Cluster ISAN×ISBN ISAN×ISBN ISAN×ISBN ISAN×ISBN ISAN×ISBN Effects (ISAN×ISBN) Fixed Random Random Random Random Notes: The dependent variable is the price discount. In all columns we control for the following variables (see Table 3 for the description of control variables): month_t, trading_m1, trading_m2, trading_m3, entry (1, 2, 3), exit (1, 2, 3), season, woodpulp, internet, the housing price, GVA, and elapsed_year. In the random effects specifications (columns 2 to 5) we additionally control for waterstone, avgsales_area, population, pop_density, urban area, universities, education, classD2, classD3, classD4, series, figure, pages, paperback. Robust t-statistics (column 1) and z-statistic (columns 2 to 5) in parentheses. The symbols ***, **, and * represent significance at the 1%, 5%, and 10% level respectively. ISAN is the Nielsen’s unique identifier of a store while ISBN is the unique identifier of a title.
36 Table 6: DiD on Local Prices – Heterogeneous Treatment Effects
All titles Bestsellers Deep-range Evergreen Top-sellers (1) (2) (3) (4) (5) WATERSTONE’S/OTTAKAR’S overlap×post×Waterstones -1.496*** -1.536** -0.560** -1.166*** -0.815 (-5.750) (-2.375) (-2.118) (-2.723) (-0.942)
overlap×post×Ottakars 0.927*** 1.540*** 0.237 1.430*** 0.542 (3.710) (2.771) (0.905) (3.364) (0.650)
BOOKS PUBLISHED POST MERGER overlap×post×released_post×just_pub 1.648** (2.245) overlap×post×released_post×non_just_pub -2.248*** (-2.816) overlap×post×released_pre -0.284 -0.097 -0.146 0.112 0.440 (-1.35) (-0.184) (-0.674) (0.310) (0.331)
STORE CLOSURES overlap×post×closed -0.363 0.183 0.146 0.166 -1.774 (-0.78) (0.182) (0.298) (0.224) (-1.139) overlap×post×non_closed -0.282 -0.091 -0.192 0.105 0.039 (-1.30) (-0.174) (-0.863) (0.285) (0.052) SPECIALIZED + GENERICS overlap×post×high comp -0.301 -0.087 -0.131 0.112 -0.630 (-1.03) (-0.133) (-0.476) (0.240) (-0.682) overlap×post×medium comp -0.345 -0.409 -0.141 0.156 -0.204 (-1.21) (-0.620) (-0.493) (0.334) (-0.230) overlap×post×low comp -0.155 0.703 -0.204 0.027 0.760 (-0.44) (0.863) (-0.552) (0.049) (0.729) SUPERMARKETS overlap×post×high comp -0.067 0.156 -0.193 0.329 -1.284 (-0.25) (0.249) (-0.705) (0.717) (-1.388) overlap×post×medium comp -0.357 -0.480 0.032 0.075 0.584 (-1.04) (-0.612) (0.096) (0.138) (0.573) overlap×post×low comp -0.530* -0.025 -0.239 -0.123 0.631 (-1.77) (-0.036) (-0.791) (-0.261) (0.707) ENTRY overlap×post×high entry -0.866** -0.365 -0.234 -0.050 -2.264 (-1.97) (-0.388) (-0.561) (-0.074) (-1.599) overlap×post×medium entry -0.023 0.424 -0.391 0.317 -0.659 (-0.07) (0.589) (-1.302) (0.607) (-0.639) overlap×post×no entry -0.275 -0.189 -0.002 0.065 0.514 (-1.11) (-0.322) (-0.008) (0.160) (0.655) EXIT overlap×post×medium exit -0.601* -0.212 -0.086 -0.338 0.035 (-1.66) (-0.269) (-0.237) (-0.607) (0.032) overlap×post×no exit -0.195 -0.012 -0.169 0.245 -0.202 (-0.87) (-0.021) (-0.742) (0.645) (-0.268) Notes: The dependent variable is the price discount. We only report coefficients for the interaction variables that represent heterogeneous treatment effects. In all columns we control for the following variables (see Table 3 for the description of control variables): month_t, trading_m1, trading_m2, trading_m3, entry (1, 2, 3), exit (1, 2, 3), season, woodpulp, internet, the housing price, GVA, and elapsed_year. In the random effects specifications (columns 2 to 5) we additionally control for waterstone, avgsales_area, population, pop_density, urban area, universities, education, classD2, classD3, classD4, series, figure, pages, paperback. Robust t-statistics (column 1) and z-statistic (columns 2 to 5) in parentheses. The symbols ***, **, and * represent significance at the 1%, 5%, and 10% level respectively.
37 Table 7: Effect of the merger on sales volumes
All titles Bestsellers Deep-range Evergreen Top-sellers (1) (2) (3) (4) (5) Overlap -0.500* -0.041* -0.309 -1.184 (-1.873) (-1.955) (-1.529) (-0.807) Post -0.495** -0.716** 0.023 -1.393*** -4.265** (-2.19) (-2.317) (0.592) (-7.875) (-2.006) overlap×post 0.245 0.152 -0.011 -0.232 1.524 (0.73) (0.405) (-0.428) (-1.359) (0.958) month_t 0.112*** 0.024** -0.004*** 0.017** -0.715*** (9.53) (2.357) (-2.939) (2.413) (-8.078) Constant 8.412* 16.590*** 1.917*** -0.581 18.220** (1.77) (10.28) (11.13) (-0.602) (2.373) Observations 172,991 37,094 58,098 56,955 20,433 R-squared 0.082 Number of id 11,833 4,544 6,909 2,445 2,930 Cluster ISAN*ISBN ISAN*ISBN ISAN*ISBN ISAN*ISBN ISAN*ISBN Time Trend YES YES YES YES YES FE ISAN*ISBN Notes: The dependent variable is monthly average volume of sales in each store. In all columns we control for the following variables (see table 3 in the paper for the description of control variables): trading_m1, trading_m2, trading_m3, entry (1, 2, 3), exit (1, 2, 3), season, woodpulp, Internet, the housing price, GVA, closed, just_pub, and elapsed_year. In the random effects specifications (columns 2 to 5) we additionally control for waterstone, avgsales_area, population, pop_density, urban area, universities, education, classD2, classD3, classD4, series, figure, pages, paperback. Robust t-statistics (column 1) and z-statistic (columns 2 to 5) in parentheses. The symbols ***, **, and * represent significance at the 1%, 5%, and 10% level respectively. ISAN is the Nielsen’s unique identifier of a store while ISBN is the unique identifier of a title.
38 Table 8: DiD on National Prices – Average Treatment Effect Competitors as control group
All titles Bestsellers Deep-range Evergreen Top-sellers (1) (2) (3) (4) (5) merged 0.225 -3.861*** 0.005 -1.483 (0.078) (-2.749) (0.003) (-0.526) post -0.732 -3.562 -0.360 -1.668 -4.795 (-0.766) (-1.538) (-0.269) (-1.262) (-0.958) post×merged 0.108 0.692 0.646 0.365 -2.751 (0.089) (0.229) (0.421) (0.176) (-0.470) constant 218.100*** 21.770 47.390** 60.62** -146.500 (3.318) (0.485) (2.165) (2.047) (-1.571)
observations 13,346 2,417 7,173 2,913 814 R-squared 0.064 number of id 400 156 270 82 98 cluster ISBN ISBN ISBN ISBN ISBN Effects (ISBN×retailer) Fixed Random Random Random Random Notes: The dependent variable is the price discount. In all columns we control for the following variables (see Table 3 for the description of control variables): month_t, season, woodpulp, ip, avg_hp, gdp_pc, just_pub, and elapsed_year. We impose two distinct time trends, one for the merging parties and one for the competitors (see Appendix 5 for a discussion). In the random effects specifications (columns 2 to 5) we additionally control for pages, series, figure, paperback, classD2, classD3, classD4. Robust t-statistics (column 1) and z- statistic (columns 2 to 5) in parentheses. The symbols ***, **, and * represent significance at the 1%, 5%, and 10% level respectively.
Table 9: DiD on National Prices – Average Treatment Effect Top-selling Titles as Control Group
Bestsellers Deep-range Evergreen (1) (2) (3) non_top -9.144*** -17.679*** -4.714** (-3.73) (-7.32) (-2.08) post -6.463 -4.812 -6.841 (-1.35) (-1.09) (-1.59) overlap×non_top 1.431 6.322 5.164 (0.36) (1.52) (1.40) Constant 53.245 19.447 35.875 (0.82) (0.70) (0.66) Observations 1,526 3,457 1,696 number of id 127 184 90 Cluster ISBN ISBN ISBN Individual Effects Random Random Random Notes: The dependent variable is the price discount. In all columns we control for the following variables (see Table 43 for the description of control variables): month_t, season, woodpulp, ip, avg_hp, gdp_pc, just_pub, elapsed_year, pages, series, figure, paperback, classD2, classD3, classD4. Robust z-statistic are reported in parentheses. The symbols ***, **, and * represent significance at the 1%, 5%, and 10% level respectively.
39 Online Appendixes Ex-post Merger Evaluation in the UK Retail Market for Books
By Luca Aguzzoni, Elena Argentesi, Lorenzo Ciari, Tomaso Duso, and Massimo Tognoni
1
Appendix 1. Selection of Stores and Areas
This Appendix describes the methodology used to select the local areas used in our empirical analysis. In its “Bookscan” database, Nielsen collects information on book sales from a wide panel of Waterstone’s and Ottakar’s stores. For the pre-merger period, this panel includes 359 Waterstone’s and Ottakar’s stores located in 203 different areas (defined at the local authorities level), of which 33 are overlap areas (as defined by the CC) and 170 are non-overlap ones. 1 To select the 60 stores for our analysis we followed an approach based on a matching methodology on observable area’s characteristics. This procedure aims at correcting for the potential sample selection bias that may affect the estimate of the treatment effects if we did not carefully choose treatment and control areas. Indeed, in non-experimental studies the assignment of subjects to the treatment and control groups cannot be considered to be random. Specifically, units receiving treatment and those excluded from treatment may differ not only in their treatment status but also in other characteristics that affect both participation and the outcome of interest. Thus the estimate of a causal effect obtained by comparing a treatment group with a non-experimental comparison group could be biased because of systematic differences between the two groups. This bias can be reduced if the comparison of outcomes is performed using treated and control groups that are as similar as possible. It might be relatively simple to assign a comparison unit based on a single observable characteristic. However, if the matching process is to be effective in mitigating the potential bias, one needs to consider a full range of factors across which the treatment and control group might differ. In our matching procedure, the degree of closeness among groups is measured by the propensity score, i.e. the probability of treatment, given a set of observed characteristics. The idea is that all relevant differences between the groups pre-treatment can be captured by observable characteristics in the data and these characteristics can be used to estimate the propensity score.2 Through this approach, a propensity score (which
1 From our selection we excluded all the shops in the London area as, although both Waterstone’s and Ottakar’s operated stores in that area, it was not considered an overlap location by the CC. 2 Once accounted for these differences, one can take assignment to treatment to have been random.
2 ranges from 0 to 1) is attached to every unit, with treatment and control groups subsequently generated and matched based on it.3 A fundamental requirement for this method is that the predicted probabilities of treatment for control and treated units must have a wide common support region, i.e. the existence of a substantial overlap between the propensity scores of control and treated units. In practice, we applied our matching methodology according to these steps: (1) Identify relevant observable characteristics; (2) Estimate the predicted probability (pscore) of assignment to treatment for all areas; and (3) Match (without replacement) each treated area with the control area that has the closest pscore. In the first step the aim is to select all the observable explanatory variables that characterize the book retailing market at the local level. These can be broadly classified into two groups: (i) factors that may impact demand and (ii) factors that may affect supply. We then estimate the predicted probability of being in an overlap by means of a logistic regression on the discrete dependent variable of treatment assignment. The results from this regression can be found in Table A1.1.
3 Note that, even if we use the word “propensity score,” this methodology does not coincide exactly with the so-called propensity score matching, because it does not take into account the fact that the propensity score is estimated when computing standard errors.
3
Table A1.1: Propensity Score Matching, estimation results Dep variable: Overlap Coeff. t-statistics Population 0.000001 (0.27) pop_density -0.0002 (-0.01) Avgsales 0.0000006 (0.09) Universities -0.00950 (-0.01) Education -0.00379 (-0.03) GVA (2004) 0.00008 (0.51) Internet (2005) -0.0419 (-0.27) house_price (2004) -0.000005 (-0.37) trading_m1 -0.0170 (-0.12) trading_m3 -0.0189 (-0.30) Scotland -0.412 (-0.33) Constant 1.659 (0.22) Observations 50
After the regression, each local area is assigned a probability of treatment. By looking at the distribution of these predicted probabilities (see Figure A1.1) we can check if the common support requirement is satisfied. We conclude that there is substantial overlap and we are then reassured that we can find a sufficient number of treated local areas with a close enough match in the control group.
4
Figure A1.1: Pscore distribution by groups and common support
0 .2 .4 .6 .8 Propensity Score
Untreated Treated: On support Treated: Off support
The selection of the treated areas was also constrained by data availability4 and out of the 33 overlap local areas we could use only 20. For each of the 20 selected local areas we found the closest match in the non-overlap areas following the PSM approach. Table A1.2 presents the final list of areas from this matching process.
4 Some shops closed or were not surveyed by Nielsen Bookscan.
5
Table A1.2: Store Matching Outcome Treated Control Location pscore store location pscore store Southend-on-Sea 0.110 W Oxford 0.111 W a) areas where we Worcester 0.269 W Nottingham 0.256 W selected only a Canterbury 0.308 W Bournemouth 0.300 W Waterstone’s store Kings Lynn 0.192 W Bath 0.189 W Milton Keynes 0.102 W Romford 0.103 W Folkestone 0.222 O Dumfries 0.222 O b) areas where we Bromley 0.033 O Barnet 0.033 O selected only a Cheltenham 0.169 O High Wycombe 0.171 O Ottakar’s store Guildford 0.162 O Barnstaple 0.159 O Harrogate 0.115 O Staines 0.115 O Aberdeen 0.469 W Bristol 0.508 W Aberdeen 0.469 O Newport 0.382 O Chelmsford 0.285 W Stirling 0.294 W Chelmsford 0.285 O Elgin 0.279 O Coventry 0.167 W Chichester 0.167 W Coventry 0.167 O Newton Abbot 0.165 O Inverness 0.230 W Winchester 0.233 W Inverness 0.230 O Loughborough 0.230 O c) areas where we Huddersfield 0.071 W Stockport 0.072 W selected both a Huddersfield 0.071 O St Albans 0.071 O Waterstone’s and a Crawley 0.203 W Derby 0.202 W Ottakar’s store Crawley 0.203 O Ashford 0.204 O Lancaster 0.198 W Wolverhampton 0.194 W Lancaster 0.198 O Andover 0.198 O Meadowhall 0.182 W Stoke On Trent 0.188 W Meadowhall 0.182 O Carlisle 0.186 O Norwich 0.309 W Leicester 0.316 W Norwich 0.309 O Aberystwyth 0.315 O Epsom 0.120 W Bedford 0.116 W Epsom 0.120 O Bishop's Stortford 0.123 O
6
A graphical representation of this selection can also be found in Figure A1.2, which shows that the matched overlap and non-overlap localities are equally spread around the UK (the only exception is for Wales, where there were no overlap areas).
Figure A1.2: Geographic distribution of Treatment and Control Areas Treated local areas Control local areas
For the selected localities we also tested the equality of means for the relevant explanatory variables and verified if the means across the two groups were not statistically different (see table A1.3).
7
Table A1.3: Test on equality of means for explanatory variables Mean t-test Variable Treated Control t p>t Pscore 0.196 0.203 -0.26 0.793 Population 180000 170000 0.48 0.633 pop_density 14.051 13.62 0.11 0.912 Universities 0.75 0.73333 0.08 0.937 Education 19.9 20.6 -0.38 0.702 Avgsales 180000 180000 -0.01 0.989 GVA_2004 16876 16630 0.22 0.824 Internet_2005 56.4 56.833 -0.39 0.702 house_price_2004 180000 190000 -0.61 0.548 trading_m1 6.15 6.3667 -0.15 0.883 trading_m3 15.15 14.767 0.11 0.916 Scotland 0.1 0.1 0 1
According to the CC, both Waterstone’s and Ottakar’s stores are almost entirely located in High Street locations. Consistently with this definition, we provide further evidence that in our overlap areas Waterstone’s and Ottakar’s indeed competed head-to- head before the merger, i.e., that the merging parties’ stores are very close to each other. In table A1.4 we list the locations in our sample together with the corresponding distance between Waterstone’s and Ottakar’s stores. With the exception of one area, stores are located within 0.2 miles (322 meters) from each other. Finally, we report the list of locations where there was a store closure in our sample in table A1.5.
8
Table A1.4: List of locations and distance between Waterstone’s and Ottakar’s stores Distance between Areas Location Waterstone’s and Ottakar’s stores (miles) Southend-on-Sea 0.1 Worcester 0.0 a) areas where we selected only Canterbury 0.0 a Waterstone’s store Kings Lynn 0.1 Milton Keynes 0.2 Folkestone 0.1 Bromley 0.1 b) areas where we selected only Cheltenham 0.1 a Ottakar’s store Guildford within 0.1 Harrogate 0.1 Aberdeen 0.2 Chelmsford 1.8 Coventry 0.2 Inverness 0.2 c) areas where we selected both Huddersfield 0.1 a Waterstone’s and a Ottakar’s Crawley 0.0 store Lancaster 0.1 Meadowhall 0.0 Norwich 0.0 Epsom 0.1 Source: “HMV Group plc and Ottakar’s plc - Proposed acquisition of Ottakar’s plc by HMV Group plc through Waterstone’s Booksellers Ltd,” 12th of May 2006, Appendix E. Distance calculations based on store postcodes.
Table A1.5: List of locations with store closures
Location Overlap Closed store’s Week closed # of stores left Brand after closure Aberdeen Y Waterstone’s 2004w39 1 Waterstone’s & 1 Ottakar’s Cheltenham Y Waterstone’s 2007w34 1 ex Ottakar’s Guildford Y 2 Waterstone’s 2006w51 & 1 ex Ottakar’s 2007 (no exact date) Harrogate Y Waterstone’s 2007w05 1 ex Ottakar’s Nottingham N Waterstone’s 2005w04 1 Waterstone’s Stoke On Trent N Waterstone’s 2004w17 1 Waterstone’s
9
Appendix 2. Selection of Book Titles
In this appendix we describe more in depth the selected book titles. We first present preliminary statistics by genre and type of binding for the selected sample in table A2.1. In table A2.2, we further list the 200 titles on which we performed the econometric analyses. Table A2.1: Number of titles by genre and type of binding Hardcover Paperback Total Fiction 6 56 62 Specialist - 20 20 Trade 16 50 66 Young 6 46 52 Total 28 172 200
Table A2.2: List of the titles included in our dataset Title Author 1 7000 Baby Names: Classic and Modern Spence, Hilary 2 Adventure of English, The Bragg, Melvyn 3 Allen Carr's Easy Way to Stop Smoking Carr, Allen 4 Amber the Orange Fairy: Rainbow Magic Meadows, Daisy 5 Angel Price, Katie 6 Angels: Miniature Editions . 7 Animal Discovery Cards: Baby Einstein S. Aigner-Clark, Julie 8 Art of Drawing Manga, The Krefta, Ben 9 Atonement McEwan, Ian 10 Bad Beginning, The: Series of Unfortunate Events Snicket, Lemony 11 Bare Bones Reichs, Kathy 12 Beginner's French: Teach Yourself Languages Carpenter, Catrine 13 Bible Code, The Drosnin, Michael 14 Blow Fly Cornwell, Patricia 15 BMA Concise Guide to Medicines and Drugs Henry, John A. 16 Body Double Gerritsen, Tess 17 Body Shape Bible, The: Forget Your Size Discover Your Shape Transform Constantine, Susanna Yourself 18 Bond Assessment Papers: Second Papers in Maths 8-9 Years: Bond Baines, Andrew & Bon
10
Assessment Papers S. 19 Broker, The Grisham, John 20 Brother's Journey, A: Surviving a Childhood of Abuse Pelzer, Richard B. 21 Brussels and Bruges: AA Citypacks Franquet, Sylvie & S 22 Castle of Wizardry: Belgariad S. Eddings, David 23 Cause of Death Cornwell, Patricia 24 Change Your Life in Seven Days McKenna, Paul 25 Chapter House Dune:(Bk. 6) :Gollancz S.F. Herbert, Frank 26 Child Called It, A Pelzer, Dave 27 Cigars of the Pharoah: The Adventures of Tintin S. Herge 28 Coast Somerville, Christop 29 Coming Out Steel, Danielle 30 Complete Beginners' Cookbook Watt, Fiona 31 Concise Colour Medical Dictionary: Oxford Paperback Reference S. Martin, Elizabeth 32 Concise Oxford Spanish Dictionary . 33 Confusion, The Stephenson, Neal 34 Contest Reilly, Matthew 35 Cranks Recipe Book, The Canter, David 36 Crucible, The: A Play in Four Acts: Penguin Modern Classics Miller, Arthur 37 Curious Incident of the Dog in the Night-time, The Haddon, Mark 38 Dark is the Moon: View from the Mirror S. Irvine, Ian 39 Dark Tower,The: D rawing of the Three (Bk. 2) King, Stephen 40 Devil's Disciples, The: The Life and Times of Hitler's Inner Circle Read, Anthony 41 Diaries 1969-1979:The Python Years Palin, Michael 42 Dr. Gillian McKeith's Ultimate Health Plan: The Diet Programme That Will McKeith, Gillian Keep You Slim for Life 43 Duck: My Thomas Story Library Awdry, W. 44 Elder Gods, The Eddings, David & Edd 45 Electrician's Guide to the Building Regulations (Approved Document P, . Electrical Safety in Dwellings) 46 Elegance Tessaro, Kathleen 47 English Grammar in Use with Answers: A Self-study Reference and Practice Murphy, Raymond Book for Intermediate Students of English 48 English Passengers Kneale, Matthew 49 Enormous Crocodile, The Dahl, Roald 50 Essential Costa Brava: AA Essential S. Kelly, Tony
11
51 Essential Teaching Skills Kyriacou, Chris 52 Face the Fire: Three Sisters Island Roberts, Nora 53 Faithless Slaughter, Karin 54 False Impression Archer, Jeffrey 55 Farm: Usborne Look and Say . 56 Filth Welsh, Irvine 57 Flat Stanley in Space Brown, Jeff 58 GCSE Double Science: Chemistry Revision Guide - Higher (Pt. 1 & 2) Parsons, Richard 59 Girls Only! All About Periods and Growing-up Stuff Parker, Victoria 60 Girls Out Late Wilson, Jacqueline 61 Girls under Pressure Wilson, Jacqueline 62 Good Night, Gorilla Rathmann, Peggy 63 Gordon Ramsay's Playing with Fire: Raw, Rare to Well Done Ramsay, Gordon 64 Great Lies to Tell Small Kids Riley, Andy 65 Harry Potter Paperback Boxed Set Rowling, J.K. 66 High Fidelity Hornby, Nick 67 High Hopes Hopkins, Billy 68 Highest Tide, The Lynch, Jim 69 Holy Bible, The: King James Version: Authorized King James Version . 70 Holy Blood and the Holy Grail, The Baigent, Michael & L 71 Horrid Henry and the Mega-mean Time Machine: (Bk. 13) : Horrid Henry Simon, Francesca 72 Horrid Henry Meets the Queen: (Bk . 12) : Horrid Henry Simon, Francesca 73 How to Boil an Egg:... And 184 Other Simple Recipes for One Arkless, Jan 74 Humble Pie Ramsay, Gordon 75 I Am Too Absolutely Small for School: Charlie & Lola Child, Lauren 76 I Know You Got Soul Clarkson, Jeremy 77 I Love Capri Jones, Belinda 78 IEE on Site Guide (BS 7671: 2001 16th Edition Wiring Regulations Including . Amendment 2: 2002) 79 "In the Night Garden" Little Library: Little Library: In the Night Garden . 80 Innocent Graves Robinson, Peter 81 Internet for Dummies, The: For Dummies S. Levine, John R. & Yo 82 Introduction to Buddhism: An Explanation of the Buddhist Way of Life Kelsang Gyatso, Gesh 83 Introductory Guide to Anatomy and Physiology, An Tucker, Louise 84 Invisible Boy, The: Magical Children S. Gardner, Sally 85 It's Not About the Bike: My Journey Back to Life Armstrong, Lance
12
86 Jasper's Beanstalk: Jasper Butterworth, Nick & 87 Jolly Postman, or, Other People's Letters, The: Or, Other People's Letters: Ahlberg, Allan & Ahl Viking Kestrel picture books 88 Jonathan Strange and Mr. Norrell Clarke, Susanna 89 Jose Mourinho: Made in Portugal - the Authorised Biography Lourenco, Luis & Mou 90 Kalahari Typing School for Men,The:No.1 Ladies' Detective Agency S. McCall Smith, Alexan 91 Kama Sutra, The: Great Sex S. Hooper, Anne 92 Krakatoa: The Day the World Exploded Winchester, Simon 93 KS1 Maths: Question Book (Pt. 1 & 2) Parsons, Richard 94 KS2 Science: SAT's Practice Papers - Levels 3-5 (bookshop) Parsons, Richard 95 KS3 Science: Revision Guide - Levels 5-7 Parsons, Richard & G 96 Last Juror, The Grisham, John 97 Last Term at Malory Towers: Malory Towers S. Blyton, Enid 98 Learning to Counsel: Develop the Skills You Need to Counsel Others Sutton, Jan & Stewar 99 Letter from America:1946-2004 Cooke, Alistair 100 Little Miss Scary: Little Miss library Hargreaves, Roger 101 Lord of the Rings, The: Return of the King (v.3) Tolkien, J. R. R. 102 Lost for Words: The Mangling and Manipulating of the English Language Humphrys, John 103 Lovely Bones, The Sebold, Alice 104 Low-Fat Meals in Minutes: "Australian Women's Weekly" Home Library Tomnay, Susan 105 Magician's Nephew, The: Chronicles of Narnia S. Lewis, C.S. 106 Mammoth Book of Extreme Science Fiction, The: Mammoth Book of S. . 107 Man Called Cash, The: The Life, Love and Faith of an American Legend Turner, Steve 108 Memoirs of a Geisha Golden, Arthur 109 Monkey Puzzle Donaldson, Julia 110 Moondust: In Search of the Men Who Fell to Earth Smith, Andrew 111 Mr. Christmas Hargreaves, Roger 112 Mr. Fussy: Mr. Men Library Hargreaves, Roger 113 Mr. Perfect Robinson, Catherine 114 Mr. Uppity: Mr. Men Library Hargreaves, Roger 115 New First Aid in English, The Maciver, Angus 116 New Pocket Dog Training Fogle, Bruce 117 New Rector, The: Tales from Turnham Malpas Shaw, Rebecca 118 Next Accident, The Gardner, Lisa 119 Nursing Calculations Gatford, J.D. & Phil 120 Nursing Practice: Hospital and Home - The Adult Alexander, Margaret
13
121 Office 2003 in Easy Steps: Colour Edition:In Easy Steps S. Copestake, Stephen 122 One Child Hayden, Torey L. 123 One Hundred Ways for a Cat to Train Its Human Haddon, Celia 124 One Hundred Years of Solitude Garcia Marquez, Gabriel 125 Other Side of the Story, The Keyes, Marian 126 Other Woman, The Green, Jane 127 Oxford English Minidictionary . 128 Oxford French Verbpack, The . 129 Oxford Reading Tree: Stage 4: Storybooks: the Storm Hunt, Roderick 130 Pale Horseman,The (Hardcover) Cornwell, Bernard 131 Pale Horseman,The (Paperback) Cornwell, Bernard 132 Peekaboo Farm!: Peekabooks S. . 133 Philip's Motoring Atlas Britain 2006:Philip's Road Atlases . 134 Philosophy: The Basics: Basics (Routledge Paperback) Warburton, Nigel 135 Picking Up the Pieces Britton, Paul 136 Pippi Longstocking Lindgren, Astrid 137 "Playboy": Bartender's Guide Mario, Thomas 138 Precious Time James, Erica 139 Pregnancy Questions and Answer Book, The Lees, Christoph & Re 140 Prince, The Machiavelli, Niccolo 141 Q Pootle 5 Butterworth, Nick 142 Quick Course in Microsoft Excel 2000:Quick Course . 143 Really Rotten Experiments: Horrible Science S. Arnold, Nick 144 Rebecca Du Maurier, Daphne 145 Recoil McNab, Andy 146 Restaurant Guide,The:2004:AA Lifestyle Guides . 147 Revenge of the Middle-aged Woman Buchan, Elizabeth 148 Rick Stein's Mediterranean Escapes Stein, Rick 149 River Cottage Meat Book, The Fearnley-Whittingsta 150 Ronnie: The Autobiography of Ronnie O'Sullivan O'Sullivan, Ronnie 151 Rottweiler, The Rendell, Ruth 152 Rough Guide to Venice, The: Rough Guide Travel Guides Buckley, Jonathan 153 RSPB Pocket Birds Elphick, Jonathan & 154 Rules of Management: The Definitive Guide to Managerial Success Templar, Richard 155 Russian Dictionary: Collins GEM . 156 Salisbury and The Plain, Amesbury:1: 50 000:OS Landranger Map .
14
157 Savage Stone Age, The: Horrible Histories S. Deary, Terry 158 Secret of Crickley Hall, The Herbert, James 159 Sexual Life of Catherine M, The Millet, Catherine 160 Sharon Osbourne Extreme: My Autobiography Osbourne, Sharon 161 Shopaholic and Sister Kinsella, Sophie 162 Silly Verse for Kids: Puffin Books Milligan, Spike 163 Silver Spoon, The . 164 Smelly Slugsy: Read-to-Me Scented Storybook: "Fifi and the Flowertots" . 165 Social Work: Themes, Issues and Critical Debates . 166 Sorceress Rees, Celia 167 South Africa: AA Explorer S. Shales, Melissa 168 Sovereign: Shardlake Sansom, C.J. 169 Spanish Verb Tenses: Practice Makes Perfect Series Richmond, Dorothy De 170 Storm of Swords, A: (1) :Song of Ice and Fire Martin, George R.R. 171 SUMO (Shut Up, Move On):The Straight Talking Guide to Creating and McGee, Paul Enjoying a Brilliant Life 172 Taking ,The Koontz, Dean 173 Tao of Pooh and Te of Piglet, The: Wisdom of Pooh S. Hoff, Benjamin 174 Thief of Time: A Discworld Novel Pratchett, Terry 175 This Little Puffin: Finger Plays and Nursery Games:Puffin Books Matterson, Elizabeth 176 Thousand Days in Venice, A: An Unexpected Romance de Blasi, Marlena 177 Thud!: Discworld Novels Pratchett, Terry 178 Time and Chance Penman, Sharon K. 179 Times Tables: Magical Skills (Level 2) :Magical skills Fidge, Louis & Broad 180 Trojan Odyssey Cussler, Clive 181 Truth, The: Discworld Novels Pratchett, Terry 182 Twelfth Card, The Deaver, Jeffery 183 Twilight Children: Three Voices No One Heard - Until Someone Listened Hayden, Torey L. 184 Twist of Gold Morpurgo, Michael 185 Twisted: Collected Stories of Jeffery Deaver Deaver, Jeffery 186 Ultimate Dinosaur Glow in the Dark Sticker Book, The: Ultimate Stickers . 187 Under Orders Francis, Dick 188 Understanding Health and Social Care: An Introductory Reader: Published in . Association with the Open University 189 Unlocking Formative Assessment: Practical Strategies for Enhancing Pupils' Clarke, Shirley Learning in the Primary Classroom
15
190 Untouchable: Alpha Force S. Ryan, Chris 191 Usborne Complete Book of Drawing: Usborne Activity Books . 192 Vesuvius Club, The: A Lucifer Box Novel Gatiss, Mark 193 Vieira: My Autobiography Vieira, Patrick 194 Wasp Factory, The Banks, Iain 195 Wedding Flowers: Over 80 Glorious Floral Designs for That Special Day Roberts, Stephen 196 Wee Free Men, The Pratchett, Terry 197 Wide Sargasso Sea: Student Edition: Penguin Modern Classics Rhys, Jean 198 Wide Window, The: Series of Unfortunate Events Snicket, Lemony 199 "York Notes on ""An Inspector Calls"": York Notes" Scicluna, John 200 Yorkshire Dales: Walks: Pathfinder Guide Conduit, Brian & Mar
16
Appendix 3. Local Price Variation
This Appendix presents some further results on the geographic scope of price competition among book retailers.
3.1. Coefficients of Variation: Time, Store, and Title.
To get a finer picture of the amount of price variability, we look at the distribution of the coefficients of variation along different dimensions and levels of aggregation. First looking at the time dimension, we compute the coefficient of variation across stores for each title on a monthly basis (figure A3.1). The amount of local price variability varies over time and within a year, but it is not straightforward to identify a clear trend.
Figure A3.1: Distribution of the monthly coefficient of variation across stores for different months
Month 2004m1 Month 2004m5 Month 2004m9 Month 2005m1 2 2 2 2 1.5 1.5 1.5 1.5 1 1 1 1 Density Density Density Density .5 .5 .5 .5 0 0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 CV CV CV CV kernel = epanechnikov, bandwidth = 0.1500 kernel = epanechnikov, bandwidth = 0.1500 kernel = epanechnikov, bandwidth = 0.1500 kernel = epanechnikov, bandwidth = 0.1500
Month 2005m5 Month 2005m9 Month 2006m1 Month 2006m5 2 2 2 2 1.5 1.5 1.5 1.5 1 1 1 1 Density Density Density Density .5 .5 .5 .5 0 0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 CV CV CV CV kernel = epanechnikov, bandwidth = 0.1500 kernel = epanechnikov, bandwidth = 0.1500 kernel = epanechnikov, bandwidth = 0.1500 kernel = epanechnikov, bandwidth = 0.1500
Month 2006m9 Month 2007m5 Month 2007m9 2 2 2 1.5 1.5 1.5 1 1 1 Density Density Density .5 .5 .5 0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 CV CV CV kernel = epanechnikov, bandwidth = 0.1500 kernel = epanechnikov, bandwidth = 0.1500 kernel = epanechnikov, bandwidth = 0.1500
Best Seller Deep range Evergreen Top
17
Second, we show the distribution of the coefficient of variation across time for the same store and for the same title (figure A3.2). The coefficients of variation are computed on a quarterly basis using monthly prices. For each combination of shop and title, we also observe quite a lot of variation over time, even within a quarter.
Figure A3.2: Distribution of the coefficient of variation across time for the same store and for the same title
Monthly prices Quarter 2004q1 Monthly prices Quarter 2004q2 Monthly prices Quarter 2004q3 Monthly prices Quarter 2004q4 1.5 1.5 1.5 1.5 1 1 1 1 .5 .5 .5 .5 Density Density Density Density 0 0 0 0
0 .5 1 1.5 2 0 .5 1 1.5 2 0 .5 1 1.5 2 0 .5 1 1.5 2 CV CV CV CV kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000
Monthly prices Quarter 2005q1 Monthly prices Quarter 2005q2 Monthly prices Quarter 2005q3 Monthly prices Quarter 2005q4 1.5 1.5 1.5 1.5 1 1 1 1 .5 .5 .5 .5 Density Density Density Density 0 0 0 0
0 .5 1 1.5 2 0 .5 1 1.5 2 0 .5 1 1.5 2 0 .5 1 1.5 2 CV CV CV CV kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000
Monthly prices Quarter 2006q1 Monthly prices Quarter 2006q2 Monthly prices Quarter 2006q3 Monthly prices Quarter 2006q4 1.5 1.5 1.5 1.5 1 1 1 1 .5 .5 .5 .5 Density Density Density Density 0 0 0 0
0 .5 1 1.5 2 0 .5 1 1.5 2 0 .5 1 1.5 2 0 .5 1 1.5 2 CV CV CV CV kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000
Kernel density estimate Kernel density estimate Kernel density estimate Kernel density estimate Monthly prices Quarter 2007q1 Monthly prices Quarter 2007q2 Monthly prices Quarter 2007q3 Monthly prices Quarter 2007q4 1.5 1.5 1.5 1.5 1 1 1 1 .5 .5 .5 .5 Density Density Density Density 0 0 0 0
0 .5 1 1.5 2 0 .5 1 1.5 2 0 .5 1 1.5 2 0 .5 1 1.5 2 CV CV CV CV kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000 kernel = epanechnikov, bandwidth = 0.2000
Best Seller Deep range Evergreen Top
3.2. Standard Deviation
In addition to the analysis of the coefficients of variation, we also calculate the standard deviation of the discounts granted by Waterstone’s and Ottakar’s across the 60 stores in our sample for each title and for each month. For each title in a given month, we consider the average percentage discount applied by each store selling that title. We then compute the standard deviation of these percentage discounts across stores. Finally, we estimate the distribution function of the standard deviation using a kernel density estimator. Figure A3.3 18
presents the distribution of the standard deviation of the discount across stores for all the titles together pre- and post-merger as well as for each of the four categories of titles. The standard deviation across stores is slightly lower (more concentrated around 0) in the post- merger period. However, the reduction in the variability after the merger is limited, which does not suggest any relevant change in the geographic scope of competition.
Figure A3.3: Distribution of monthly standard deviation: pre vs. post-merger
All titles Monthly prices .1 .08 .06 Density .04 .02 0
0 10 20 30 40 stdv of monthly % discount
Pre-Merger Post-Merger
kernel = epanechnikov, bandwidth = 1.5000
Best sellers Deep Range Monthly prices Monthly prices .15 .15 .1 .1 Density Density .05 .05 0 0
0 10 20 30 40 0 10 20 30 40 stdv of monthly % discount stdv of monthly % discount kernel = epanechnikov, bandwidth = 1.0000 kernel = epanechnikov, bandwidth = 1.0000
Evergreen Top Monthly prices Monthly prices .15 .15 .1 .1 Density Density .05 .05 0 0
0 10 20 30 40 0 10 20 30 40 stdv of monthly % discount stdv of monthly % discount kernel = epanechnikov, bandwidth = 1.0000 kernel = epanechnikov, bandwidth = 1.0000
Pre-Merger Post-Merger
As discussed in the paper, the discount variability differs across categories. Top- selling titles and, to a lesser extent, evergreen titles have a high standard deviation, while the 19
discount variability of bestsellers and deep-range titles appears to be lower and concentrated around 0. This latter result suggests that for these titles price competition mainly occurs at national level. However, we cannot say, in particular for deep-range titles, whether this low variability was due to a strict application of a centrally set pricing policy, or to the fact that local conditions did not vary much (for example, because the demand for deep-range titles was scarcely elastic over the entire nation).
3.3. Percentiles of the Discount Distribution
To further investigate the issue of local price variation, we also examined the percentiles of the discount distribution. For each title in each month, we derived the percentiles of this distribution and analyzed them graphically. A higher vertical difference between percentiles would suggest higher dispersions of the discount across stores. In figure A3.4 we plot a relatively narrow interval (percentiles 25th 50th and 75th), whereas in figure A3.5 we plot a larger interval (percentiles 10th 50th and 90th).
Figure A3.4: Percentiles of the discount distribution (25th, 50th, and 75th)
Best Seller Deep Range Merger Date Merger Date 40 40 30 30 20 20 10 10 0 0
2004m1 2005m6 2006m10 2007m12 2004m1 2005m6 2006m10 2007m12
Evergreen Top Merger Date Merger Date 40 40 30 30 20 20 10 10 0 0
2004m1 2005m6 2006m10 2007m12 2004m1 2005m6 2006m10 2007m12
25th pct 50th pct 75th pct
20
Figure A3.5: Percentiles of the discount distribution (10th, 50th, and 90th)
Best Seller Deep Range Merger Date Merger Date 40 40 30 30 20 20 10 10 0 0
2004m1 2005m6 2006m10 2007m12 2004m1 2005m6 2006m10 2007m12
Evergreen Top Merger Date Merger Date 40 40 30 30 20 20 10 10 0 0
2004m1 2005m6 2006m10 2007m12 2004m1 2005m6 2006m10 2007m12
10th pct 50th pct 90th pct
The percentiles analysis confirms the previous results. We observe a high variability for evergreen and top-selling titles, a lower one for best-sellers and a very low one for deep range titles.
3.4. Waterstone’s vs. Ottakar’s
Finally, we also verified whether there was any difference in the pricing policies adopted by Waterstone’s and Ottakar’s before the merger. This was done in order to check the opinions expressed by some market participants5 who claimed that Ottakar’s tended to have a more local-oriented pricing policy. Hence, we computed the discount variability across Waterstone’s stores before and after the merger and compared it with the same figures for Ottakar’s. In figure A3.6 we plot the distribution of the discount standard deviations across Waterstone’s and Ottakar’s stores before the merger.
5 These opinions were expressed to both the CC during its inquiry and to us in the responses of our questionnaires. 21
Figure A3.6: Distribution of monthly standard deviation before merger: Waterstone’s vs Ottakar’s
All Monthly prices .1 .08 .06 Density .04 .02 0
0 10 20 30 40 STDV
Waterstone's Ottakar's
kernel = epanechnikov, bandwidth = 1.5000
Best Seller Deep Range Monthly prices Monthly prices .15 .15 .1 .1 Density Density .05 .05 0 0
0 10 20 30 40 0 10 20 30 40 STDV STDV kernel = epanechnikov, bandwidth = 1.5000 kernel = epanechnikov, bandwidth = 1.5000
Evergreen Top Monthly prices Monthly prices .15 .15 .1 .1 Density Density .05 .05 0 0
0 10 20 30 40 0 10 20 30 40 STDV STDV kernel = epanechnikov, bandwidth = 1.5000 kernel = epanechnikov, bandwidth = 1.5000
Waterstone's Ottakar's
The graphical inspection shows no significant difference between the merged parties, suggesting that before the merger the extent to which Waterstone’s and Ottakar’s adopted
22
local pricing was similar. As expected, this holds all the more for the discounts applied once the merger was consummated (figure A3.7).
Figure A3.7: Distribution of monthly standard deviation after the merger: Waterstone’s vs Ottakar’s
All Monthly prices .1 .05 Density 0
0 10 20 30 40 STDV
Waterstone's Ottakar's
kernel = epanechnikov, bandwidth = 1.5000
Best Seller Deep Range Monthly prices Monthly prices .15 .15 .1 .1 Density Density .05 .05 0 0
0 10 20 30 40 0 10 20 30 40 STDV STDV kernel = epanechnikov, bandwidth = 1.5000 kernel = epanechnikov, bandwidth = 1.5000
Evergreen Top Monthly prices Monthly prices .15 .15 .1 .1 Density Density .05 .05 0 0
0 10 20 30 40 0 10 20 30 40 STDV STDV kernel = epanechnikov, bandwidth = 1.5000 kernel = epanechnikov, bandwidth = 1.5000
Waterstone's Ottakar's
23
Appendix 4. Local Analysis: Additional Regressions & Robustness Checks
This Appendix presents several specifications of our main regressions to deal with different econometric issue. These constitute additional robustness checks for the results reported in the paper.
4.1. Comparison of the estimated effects using title/store fixed-effects and title/store characteristics
We compare the estimated price effects of the merger under two different specifications of the individual effects. In the first, we use product fixed-effects to control for unobserved heterogeneity, while in the second we include a set of time invariant product characteristics that may affect a title’s price and estimate a random-effects model. All specifications are estimated on the set of titles that are in the same category both before and after the merger, because it is only for these titles that the price effects are identified in the presence of product fixed-effects. We show this comparison for all categories but top-sellers, since very few titles are in this category both in the pre- merger and in the post-merger period, and therefore the fixed-effects specification is not feasible. Table A4.1 shows the results for both specifications. The fact that the estimate of treatment effect is similar under both specifications suggests that unobservable product characteristics do not result in bias of the estimator of the random-effect regression.6
6 The category that exhibits the larger difference in the two coefficients is the bestsellers’. However these results have to be taken more cautiously because only a small proportion of titles (27 out of 62) belong to this category both in the pre-merger and in the post-merger period. 24
Table A4.1: Comparison of estimated effects using fixed-effects and product characteristics
Deep-range Evergreen Bestsellers (1) (2) (3) post×overlap -0.245 -0.193 0.039 0.090 -0.701 -0.127 (-1.02) (-0.84) (0.10) (0.24) (-1.11) (-0.22) Title/store fixed-effects YES NO YES NO YES NO Title and Store Characteristics NO YES NO YES NO YES Observations 53,577 53,577 57,974 57,974 16,912 16,912 Notes: Title and store characteristics used are: waterstone, avgsales_area, population, pop_density, urban area, universities, education, classD2, classD3, classD4, series, figure, pages, paperback. Robust t-statistics (for the fixed-effects specifications) and z-statistic (for random-effects specifications) in parentheses.
4.2. Specification with Monthly Dummies instead of a Time Trend
We estimate a version of our main specification that includes time period dummies instead of a post-merger dummy. The effect of the merger is then captured by the interactions of the time period dummies and the overlap dummy, whose coefficients and t-statistics are reported in table A4.2. The results of this alternative specification are not qualitatively different from the original ones. In particular, the coefficients of the interaction dummies are almost never significantly different from 0.
25
Table A4.2: Interactions of month dummies with overlap dummy All titles Bestsellers Deep-range Evergreen Top-sellers (1) (2) (3) (4) (5) Coeff. t-stat Coeff. t-stat Coeff. t-stat Coeff. t-stat Coeff. t-stat 2006m6×overlap 0.258 0.55 0.022 0.017 0.149 0.311 1.474* 1.811 0.503 0.305 2006m7×overlap 0.154 0.32 1.092 0.829 0.016 0.034 0.829 0.962 -0.217 -0.112 2006m8×overlap -0.257 -0.58 0.41 0.321 -0.145 -0.264 0.129 0.171 0.68 0.392 2006m9×overlap -0.772* -1.72 -0.723 -0.533 -0.427 -0.819 -0.341 -0.47 0.965 0.55 2006m10×overlap -1.111** -2.51 -0.955 -0.738 -1.020** -2.022 -1.092 -1.326 1.654 0.876 2006m11×overlap -1.196* -2.69 0.474 0.41 -1.174** -1.987 -1.03 -1.408 -0.248 -0.15 2006m12×overlap 0.156 0.37 0.602 0.625 -0.025 -0.053 0.781 1.111 0.535 0.315 2007m1×overlap -0.23 -0.6 -0.040 -0.043 0.438 1.115 -0.147 -0.224 1.435 0.926 2007m2×overlap 0.028 0.06 0.749 0.559 -0.073 -0.208 0.503 0.735 2007m3×overlap -0.041 -0.1 1.097 1.029 0.118 0.336 0.267 0.362 2007m4×overlap -0.251 -0.59 -0.821 -0.814 0.18 0.431 0.519 0.654 2007m5×overlap 0.047 0.11 0.069 0.065 0.347 0.952 0.462 0.629 2007m6×overlap 0.012 0.03 1.021 0.879 -0.091 -0.209 0.416 0.551 0.529 0.25 2007m7×overlap 0.126 0.29 1.39 1.228 -0.366 -0.768 0.873 1.148 1.522 0.761 2007m8×overlap -0.475 -1.23 -0.138 -0.141 -0.513 -1.105 0.653 0.865 0.218 0.151 2007m9×overlap 0.305 0.78 0.747 0.847 0.582 1.087 1.259 1.622 0.389 0.285 2007m10×overlap 0 0 0.927 1.051 0.013 0.035 0.308 0.4 2.788** 2.246 2007m11×overlap -0.033 -0.08 0.67 0.901 0.11 0.2 0.517 0.722 0.905 1.014 2007m12×overlap 0.431 0.92 1.597 2.118 -0.0871 -0.305 0.967 1.691 0.808 0.851 Observations 204,013 43,460 68,773 68,328 23,041 Number of id 11,849 4,559 7,064 2,446 2,939 R-squared 0.077 Cluster ISAN×ISBN ISAN×ISBN ISAN×ISBN ISAN×ISBN ISAN×ISBN Effects Fixed Random Random Random Random Notes: The dependent variable is the price discount. In all columns we control for the following variables (see table 3 in the paper for the description of control variables): overlap, month dummies, trading_m1, trading_m2, trading_m3, entry (1, 2, 3), exit (1, 2, 3), season, woodpulp, Internet, the housing price, GVA, closed, just_pub, and elapsed_year. In the random effects specifications (columns 2 to 5) we additionally control for waterstone, avgsales_area, population, pop_density, urban area, universities, education, classD2, classD3, classD4, series, figure, pages, paperback. The symbols ***, **, and * represent significance at the 1%, 5%, and 10% level respectively. ISAN is the Nielsen’s unique identifier of a store while ISBN is the unique identifier of a title.
This specification also allows us to verify that there is no differential trend between treatment and control groups prior to the merger, which is crucial for the estimation of our treatment effect. In order to do this, we plot the coefficients on the interaction terms of table A4.2 (together with their confidence intervals) against the corresponding calendar date (figures A4.1 and A4.2) and then run a formal test of the null that there are no differential trends. These coefficients bounce around zero over the whole sample period, and in particular prior to the merger. This is what matters for the validity of our DiD estimator.
26
Figure A4.1: Coefficients of interactions between month dummies and overlap: All titles 5
3
1
-1
-3
-5 Mar-04 Sep-04 Apr-05 Oct-05 May-06 Nov-06 Jun-07
95% CI_low Coeff. 95% CI_high
27
Figure A4.2: Coefficients of interactions between month dummies and overlap: By category
(a) Best sellers (b) Deep-range 5 5
3 3
1 1
-1 -1
-3 -3
-5 -5
Mar-04 Jul-05 Nov-06 Mar-0495% Jul-05 CI_low Nov-06Coeff. 95% CI_low Coeff. 95% CI_high 95% CI_high
(c) Evergreen (d) Top-sellers7 5 5
3 3
1 1
-1 -1
-3 -3
-5 -5 Mar-04 Jul-05 Nov-06 Mar-04 Jul-05 Nov-06
95% CI_low Coeff. 95% CI_high 95% CI_low Coeff. 95% CI_high
To further check our common trend assumption, we explicitly test the hypothesis that the sum of the interaction coefficients is not significantly different from 0, both in the pre- merger and in the post-merger periods. The test results in table A4.3 show that the null- hypotheses cannot be rejected for any of the categories.
7 Note that, due to data limitations, we exclude data from the first four months of 2007 (see section 4.2 in the paper). 28
Table A4.3: Test on sum of interaction coefficients (H0: sum=0) All titles Bestsellers Deep-range Evergreen Top-sellers Pre Post Pre Post Pre Post Pre Post Pre Post Ȥ2(1) 0 0.31 0.05 0.3 0.06 0.33 0.83 0.86 0.01 0.53 Prob> Ȥ2 0.9851 0.5766 0.8227 0.5854 0.8059 0.5636 0.3616 0.3547 0.9155 0.4648
4.3. Specification with Standard Errors Clustered at the Store Level
Since there might be correlation among the prices of different titles at the store level, we estimate an alternative specification where standard errors are clustered at the store level. In table A4.4, we present both the original specification with standard errors clustered at the title/store level (columns 1-5) and the new specification with standard errors clustered only at the store level only (columns 6-10). Clustering at the store level leads to an increase in the standard errors but it does not affect the significance of the coefficients of interest.
29
Table A4.4: Comparison of estimates with different clustering of standard errors
Top- All titles Bestsellers Deep-range Evergreen Top-sellers All titles Bestsellers Deep-range Evergreen sellers (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) overlap 0.369 0.121 0.041 0.503 0.369 0.121 0.041 0.503** (0.974) (0.540) (0.122) (1.085) (1.433) (0.825) (0.172) (1.967) post 1.361*** 0.987* 2.363*** -0.387 -3.469*** 1.361*** 0.987** 2.363*** -0.387 -3.469*** (5.94) (1.881) (9.789) (-1.015) (-3.469) (4.88) (2.429) (10.86) (-1.023) (-4.554) overlap×post -0.290 -0.061 -0.148 0.112 -0.147 -0.290 -0.061 -0.148 0.112 -0.147 (-1.38) (-0.119) (-0.684) (0.310) (-0.202) (-0.89) (-0.117) (-0.661) (0.289) (-0.252) month_t -0.117*** -0.221*** -0.116*** -0.009 -0.008 -0.117*** -0.221*** -0.116*** -0.009 -0.008 (-10.77) (-13.37) (-14.41) (-0.572) (-0.329) (-10.17) (-17.14) (-18.71) (-0.877) (-0.432) Constant 18.701*** 10.890*** 14.85*** 5.596*** 11.140*** 18.701*** 10.890*** 14.850*** 5.596*** 11.140*** (5.77) (4.617) (11.23) (2.844) (3.371) (3.99) (4.470) (12.50) (4.090) (4.031)
Observations 172,991 37,094 58,098 56,955 20,433 172,991 37,094 58,098 56,955 20,433 R-squared 0.061 0.061 Number of id 11,833 4,544 6,909 2,445 2,930 11,833 4,544 6,909 2,445 2,930 Cluster ISAN*ISBN ISAN*ISBN ISAN*ISBN ISAN*ISBN ISAN*ISBN ISAN ISAN ISAN ISAN ISAN Time Trend YES YES YES YES YES YES YES YES YES YES FE ISAN*ISBN ISAN*ISBN Notes: The dependent variable is the price discount. In columns (1) to (5) standard errors are clustered at the level of title and store. In columns (6) to (10) standard errors are clustered at the level of store. In all columns we control for the following variables (see table 3 in the paper for the description of control variables): trading_m1, trading_m2, trading_m3, entry (1, 2, 3), exit (1, 2, 3), season, woodpulp, Internet, the housing price, GVA, closed, just_pub, and elapsed_year. In the random effects specifications (columns 2 to 5) we additionally control for waterstone, avgsales_area, population, pop_density, urban area, universities, education, classD2, classD3, classD4, series, figure, pages, paperback. Robust t-statistics (column 1) and z-statistic (columns 2 to 5) in parentheses. The symbols ***, **, and * represent significance at the 1%, 5%, and 10% level respectively. ISAN is the Nielsen’s unique identifier of a store while ISBN is the unique identifier of a title.
30
4.4. Effect of the Merger Including the Merger Window
Instead of dropping data from the window around the merger, we allow the treatment effect to vary within this time period by including an interaction term between treatment variable and a dummy for this time period. We include all data within the 6 month merger window and split this merger window into two periods, one before the merger date and one after. Therefore we have 4 reference periods: 1) up to three months before the merger; 2) the three months before the merger; 3) the three months after the merger; and 4) from four months after the merger. We then interact the overlap dummy with each of these four sub-periods, in order to check whether the price reaction to the merger was different in overlap with respect to non-overlap areas within each of these periods. Table A4.5 shows that, consistently with our previous results, not even within the merger window does the merger seem impact prices.
31
Table A4.5: Effect of the merger including the merger window (± 3 months)
All titles Bestsellers Deep-range Evergreen Top-sellers (1) (2) (3) (4) (5) Overlap 0.354 0.105 0.030 0.557 (0.964) (0.485) (0.092) (1.195) Post 0.847*** 1.352*** 1.558*** -0.950*** -2.339** (3.96) (2.686) (6.895) (-2.690) (-2.543) Window pre merger (3m) 1.548*** 2.686*** 1.136*** 0.250 5.793*** (8.06) (5.132) (5.366) (0.897) (7.709) Window post merger (3m) 1.461*** 1.402*** 1.274*** 0.213 2.311** (6.48) (2.626) (5.092) (0.589) (2.464) overlap×post -0.323 0.017 -0.173 0.111 -0.429 (-1.58) (0.033) (-0.829) (0.314) (-0.581) Window post merger×overlap -0.052 0.333 -0.035 0.586 -1.172 (-0.19) (0.538) (-0.106) (1.222) (-1.097) Window pre merger×overlap -0.262 0.156 -0.440 0.003 0.118 (-1.05) (0.222) (-1.560) (0.008) (0.123) Constant 18.563*** 11.200*** 10.510*** 3.498* 13.420*** (5.91) (4.873) (9.155) (1.910) (4.143) month_t -0.101*** -0.228*** -0.107*** -0.005 -0.019 (-9.43) (-13.86) (-13.52) (-0.346) (-0.831) Observations 204,013 43,460 68,773 68,328 23,041 R-squared 0.066 Number of id 11,849 4,559 7,064 2,446 2,939 Cluster ISAN*ISBN ISAN*ISBN ISAN*ISBN ISAN*ISBN ISAN*ISBN Time Trend YES YES YES YES YES FE ISAN*ISBN Notes: The dependent variable is the price discount. In all columns we control for the following variables (see table 3 in the paper for the description of control variables): trading_m1, trading_m2, trading_m3, entry (1, 2, 3), exit (1, 2, 3), season, woodpulp, Internet, the housing price, GVA, closed, just_pub, and elapsed_year. In the random effects specifications (columns 2 to 5) we additionally control for waterstone, avgsales_area, population, pop_density, urban area, universities, education, classD2, classD3, classD4, series, figure, pages, paperback. Robust t-statistics (column 1) and z-statistic (columns 2 to 5) in parentheses. The symbols ***, **, and * represent significance at the 1%, 5%, and 10% level respectively. ISAN is the Nielsen’s unique identifier of a store while ISBN is the unique identifier of a title.
4.5. Impact of Local Competition Variables
Additional evidence on the extent of local price variation can be inferred from the role that variables measuring local competition play in determining prices. In the main regression on local prices (table 5 in the paper), we also control for variables that measure the intensity of local competition, such as the number of competing retailers that are active in the area and the number of competitors that entered/exited the market in the previous three months. For all these dimensions we can use disaggregated information on three types of competitors: other
32
specialized stores, non-specialized stores, and supermarkets. Since we did not report these coefficients in the paper, we show them in table A4.6.
Table A4.6: Impact of local competition variables
All titles Bestsellers Deep-range Evergreen Top-sellers (1) (2) (3) (4) (5) # specialist retailers -0.311*** -0.108 -0.103 -0.183 -0.403** (-2.58) (-0.674) (-1.134) (-1.261) (-2.211) # non-specialist retailers -0.018 0.033 0.034 -0.066** -0.019 (-0.69) (0.914) (1.523) (-1.994) (-0.425) # supermarkets and other retailers 0.293*** 0.047 -0.030 0.077* 0.093* (5.72) (1.027) (-1.101) (1.919) (1.692) Entry specialist 0.550** 0.631 0.206 0.795** 0.609 (2.54) (1.423) (0.899) (2.008) (1.019) Entry non-specialist 0.522*** 0.213 0.337*** 0.665*** 0.661** (5.78) (1.039) (3.132) (3.841) (2.430) Entry supermarkets and other retailers -0.306*** -0.797*** -0.196 0.135 -0.328 (-2.80) (-3.674) (-1.592) (0.592) (-1.180) Exit specialist -0.765*** -0.896** -0.153 -0.732 -0.828 (-3.22) (-1.972) (-0.641) (-1.379) (-1.320) Exit non-specialist -1.080*** -0.518** -0.402*** -1.217*** -2.680*** (-9.12) (-2.233) (-3.208) (-4.825) (-8.549) Exit supermarkets and other retailers -0.229* -0.395 -0.134 -0.479* -0.394 (-1.70) (-1.415) (-0.791) (-1.906) (-0.769)
Observations 172,991 37,094 58,098 56,955 20,433 Number of id 11,833 4,544 6,909 2,445 2,930 R-squared 0.061 Cluster ISAN×ISBN ISAN×ISBN ISAN×ISBN ISAN×ISBN ISAN×ISBN Effects (ISAN×ISBN) Fixed Random Random Random Random Notes: The dependent variable is the price discount. The specification used is the same of table 5 in the paper. In all columns we control for the following variables (see table 3 for the description of control variables): trading_m1, trading_m2, trading_m3, entry (1, 2, 3), exit (1, 2, 3), season, woodpulp, Internet, the housing price, GVA, closed, just_pub, and elapsed_year. In the random effects specifications (columns 2 to 5) we additionally control for waterstone, avgsales_area, population, pop_density, urban area, universities, education, classD2, classD3, classD4, series, figure, pages, paperback. Robust t-statistics (column 1) and z-statistic (columns 2 to 5) in parentheses. The symbols ***, **, and * represent significance at the 1%, 5%, and 10% level respectively. ISAN is the Nielsen’s unique identifier of a store while ISBN is the unique identifier of a title.
33
Appendix 5. National Analysis: Descriptive Statistics & Additional Results
In this appendix we report additional analyses and results on the effect of the merger on prices at the national level. We first provide descriptive statistics for the set of variables used in the national analysis in table A5.1.
Table A5.1: Descriptive statistics of control variables for national analysis
Variable Obs. Mean Std. Dev. post 24236 0.46 0.50 discount 23461 11.27 11.56 month_t 24236 26.21 13.47 season 24236 0.17 0.37 waterstone 24236 0.33 0.47 classD1 24236 0.29 0.45 classD2 24236 0.11 0.31 classD3 24236 0.31 0.46 classD4 24236 0.29 0.45 series 24236 0.32 0.47 Figure 24236 0.53 0.50 Pages 24236 313.23 265.84 elapsed_year 24236 3.81 3.89 just_pub 24236 0.04 0.20 paperback 24236 0.88 0.32 woodpulp 24236 5527.23 527.41 Internet 24236 56.52 4.18 house_price 24236 199573.00 15123.11 GDP per capita 24236 0.021 0.001
34
5.1. Competitors’ Titles as Control Group
Similar to the local level analysis, we start by plotting and graphically comparing the average discounts applied by the merging parties and the competitors (figure A5.1).
Figure A5.1: Distribution of monthly national discounts: merging parties vs competitors
Best Seller Deep Range Merger window Merger window 50 50 40 40 30 30 20 20 10 10 0 0 (% discount over RRP) over discount (% RRP) over discount (% 2004m1 2005m7 2006m52007m2 2007m12 2004m1 2005m7 2006m52007m2 2007m12
Evergreen Top Merger window Merger window 50 50 40 40 30 30 20 20 10 10 0 0 (% discount over RRP) over discount (% RRP) over discount (% 2004m1 2005m7 2006m52007m2 2007m12 2004m1 2005m7 2006m52007m2 2007m12
Competitors mergPart
The discount patterns of the merging parties and competitors diverge over time. The former decreased their discounts, while the latter increased them. This appears to hold for all categories, except for top-selling titles, for which no clear trend can be identified either for the merging parties or their competitors. The diverging trend seems to start indicatively around the beginning of 2005, which is well before the merger was consummated. This may hardly be the result of the merger because it would imply that the parties started acting as a single entity one and a half year before the CC’s decision.8 We consider it more likely that the observed trends are the results of structural changes in the supply-side of the market. The “competitors” group contains a wide set of retailers, ranging from specialist and non- specialist chains to supermarkets and Internet retailers. According to the data provided by the
8 Some studies find evidence of anticipatory price increases before the parties were legally granted permission to merge (e.g. Weinberg, 2008). However, even if one considers the first announcement of Waterstone’s bid for Ottakar’s in August 2005 as the date when the parties started acting as a single entity, the diverging trend began some 8 months earlier. This seems to rule out the hypothesis that the merger triggered the negative trend in the discounts. 35
Booksellers Association,9 the market shares held by supermarkets and Internet retailers have continually increased over the past years, as shown in figure A5.2.
Figure A5.2: Retailing book market: market share by volumes
ChartTitle 45
40
35
30 LargeChains 25 Shares Supermarkets 20 Internet Market 15 Independents
10
5
0 2004 2005 2006 2007 2008
Source: http://www.booksellers.org.uk/
This observation, in combination with the fact that supermarkets and Internet retailers tend to apply larger discounts than specialist and non-specialist book retailers,10 may explain, at least partially, the increase in the average discounts for the “competitors” category.11 In other words, the apparent increasing trend may simply be due to a change in the composition of the group of competitors, where the increasing weight of supermarkets and Internet retailers drives the observed pattern of the average discount. The fact that the discounts of the merging parties and the competitors show two diverging trends, which started before the merger, poses concerns on the validity of the DiD approach. Indeed, the common trend assumption that lies at the very heart of the DiD methodology seems to be violated in this case. We try to address this issue by imposing two
9 Booksellers Association website: http://www.booksellers.org.uk/Industry-Info/Industry-Reports/Book- Industry-Statistics/UK-Book-Sales---Retail-1999-2008.aspx, visited October 2010. 10 Anecdotal evidence on this is provided by a survey of market participants that we conducted for the CC (see Aguzzoni et al., 2011). Clay et al. (2002) instead find that, in the US market, online and physical stores have similar prices, although online prices are characterized by a higher dispersion. 11 This issue could have been addressed by splitting the data by retail channel and using only the large chains and independent shops channels as a control group. Unfortunately, Nielsen could not provide us with the data by retail channel due to confidentiality concerns. 36
distinct trends, one for the merging parties (month_t_merged) and one for the competitors (month_t_comp), so as to isolate the effect of the trends from that of the merger.12
5.2. Top-selling Titles as Control Group Figure A5.3 shows that the size of the discount is the largest for the top-selling category. In figure A5.3 we can also observe that, although the pre-merger price patterns in the control and the treatment groups vary across categories, they do not show diverging trends, which suggests that the assumption of common trends may be reasonable in this case.
Figure A5.3: Average discounts by category – the merging parties
Merger window 50 40 30 20 (% discount(% over RRP) 10 0
2004m1 2005m7 2006m5 2006m12 2007m12
DR TOP EG BS
5.3. Comparison of estimated effects using title/store fixed-effects and title/store characteristics
Similar to the analysis presented in appendix 4 at the local level, we present a comparison of the estimated price effects of the merger under two different specifications also at the
12 The resulting estimates might be, nonetheless, biased as the linear trends may not be able to fully capture the different dynamic of discounts for the merged parties and for the competitors. 37
national level with competitors as the control group. In the first specification, we use product fixed-effects to control for unobserved heterogeneity, while in the second we include a set of time invariant product characteristics that may affect a title’s price and estimate a random- effects model. All specifications are estimated on the set of titles that are in the same category both before and after the merger, because it is only for these titles that the price effects are identified in the presence of product fixed-effects. We show this comparison for all categories but top-sellers, since very few titles are in this category both in the pre- merger and in the post-merger period, and therefore the fixed-effects specification is not feasible. Table A5.2 shows that also at the national level, the estimated coefficients are similar under the two different specifications.
Table A5.2: Comparison of estimated effects using fixed-effects and product characteristics Deep-range Evergreen Bestsellers post×merged 0.814 0.798 0.365 0.386 2.301 2.375 (0.522) (0.510) (0.176) (0.186) (0.781) (0.801) Title fixed-effects YES NO YES NO YES NO Title characteristics NO YES NO YES NO YES Observations 6,422 6,422 2,913 2,913 1,041 1,041 Notes: The title characteristics used are: pages, series, figure, paperback, classD2, classD3, classD4. Robust t-statistics (for fixed-effects specifications) and z-statistic (for random-effects specifications) in parentheses.
5.4. Heterogeneous Treatment Effects
To mimic the heterogeneous treatment effects regressions performed at the local level, we also estimate similar specifications with the aggregated national data. In particular, we look at whether the national discounts of Waterstone’s and Ottakar’s in the different categories converged after the merger when compared to their competitors.13 Moreover, we analyze whether the timing of the discounts on books published post-merger changed at the national level as well. In both cases, the results reported in table A5.3 suggest that the effects are generally not significant. The apparent lack of significant heterogeneous treatment effects at the national level (in contrast to what we observe at the local level) is not necessarily worrying. Indeed, this additional analysis should identify complementary nation-wide effects
13 We also performed the same exercise using top sellers as control group, and the results are qualitatively similar. 38
that are derived through a different empirical analysis, performed with different empirical strategies –in particular with different samples and control groups. Overall, our findings suggest that the merger did not affect aggregate national prices, while it affects to some extent competition at the local level by altering the way the merging parties set prices in the overlap areas, although the average effect is insignificant also at this level of aggregation.
Table A5.3: DiD on National Prices – Heterogeneous Treatment Effects Competitors as control group
All titles Bestsellers Deep-range Evergreen Top-sellers (1) (2) (3) (4) (5) WATERSTONE’S/OTTAKAR’S post×merged×Waterstones 0.236 0.243 0.685 0.452 -1.301 (0.210) (0.088) (0.453) (0.248) (-0.239) post×merged×Ottakar 0.916 1.848 0.840 1.813 1.592 (0.816) (0.712) (0.558) (0.994) (0.344) BOOKS PUBLISHED POST MERGER post×merged×released_post×just_pub 8.571* (1.880) post×merged×released_post×non_just_pub 3.353 (0.696) post×merged×released_pre -5.850 (-1.030)
Notes: The dependent variable is the price discount. We only report coefficients for the interaction variables that represent heterogeneous treatment effects. In all columns we control for the following variables (see Table 3 for the description of control variables): season, woodpulp, ip, avg_hp, gdp_pc, just_pub, and elapsed_year. We impose two distinct time trends for the merging parties and for competitors. In the random effects specifications (columns 2 to 5) we additionally control for pages, series, figure, paperback, classD2, classD3, classD4. Robust t-statistics (column 1) and z-statistic (columns 2 to 5) in parentheses. The symbols ***, **, and * represent significance at the 1%, 5%, and 10% level respectively.
39